Dataset Viewer
report_id
stringlengths 36
36
| report_status
stringclasses 1
value | report_types
stringclasses 4
values | reporter_id
stringlengths 7
24
| submission_timestamp
stringdate 2025-06-02 13:35:00
2025-09-18 21:14:31
| form_data
stringlengths 290
4.42k
| machine_readable
stringlengths 379
10.4k
|
|---|---|---|---|---|---|---|
ce38606c-85d7-4bc9-a49d-aa5fd59c36e7
|
Unknown
|
["Real-World Incidents", "Malign Actor", "Security Incident Report"]
|
afasdfaw
|
2025-06-02T17:15:55.937908
|
{"Submitter Relationship": "Independent observer", "Submitter_Relationship_Other": "", "Incident Location(s)": "adfaf", "Harm Narrative": "awdsfadf", "Tactic Select": ["Execution"], "Tactic_Select_Other": "", "Detection": ["Monitoring"], "Detection_Other": "", "Disclosure Intent": "No", "Disclosure Timeline": null, "Disclosure Channels": [], "Disclosure_Channels_Other": "", "Embargo Request": "", "Report Types": ["Real-World Incidents", "Malign Actor", "Security Incident Report"], "Reporter ID": "afasdfaw", "Session ID": "awsdfawdf", "Flaw Timestamp Start": "2025-06-02", "Systems": ["o3-mini"], "Incident Description": "**Detailed Description:**\ndasf\n\n**Undesirable Outputs/Effects/Impacts:**\nadswcf\n\n**Reproduction Steps:**\naefadef\n\n**Systematic Evidence:**\nawfawdf", "Incident Description - Detailed": "dasf", "Incident Description - Outputs": "adswcf", "Incident Description - Reproduction": "aefadef", "Incident Description - Systematic": "awfawdf", "Policy Violation": "awfasdf", "Prevalence": "Rare", "Severity": "Medium", "Impacts": ["Psychological"], "Impacts_Other": "", "Impacted Stakeholder(s)": ["Developers"], "Risk Source(s)": ["Implementation error"], "Context Info": "awdfawsf", "Report ID": "ce38606c-85d7-4bc9-a49d-aa5fd59c36e7", "Submission Timestamp": "2025-06-02T17:15:54.811029"}
|
{"@context": "https://schema.org", "@type": "AIFlawReport", "reportId": "ce38606c-85d7-4bc9-a49d-aa5fd59c36e7", "dateCreated": "2025-06-02T17:15:54.923516", "reportStatus": null, "reportTypes": ["Real-World Incidents", "Malign Actor", "Security Incident Report"], "basicInformation": {"reporterId": "afasdfaw", "sessionId": "awsdfawdf", "flawTimestampStart": "2025-06-02", "flawTimestampEnd": null, "systems": ["o3-mini"]}, "commonFields": {"contextInfo": "awdfawsf", "flawDescription": null, "policyViolation": "awfasdf", "severity": "Medium", "prevalence": "Rare", "impacts": ["Psychological"], "impactedStakeholders": ["Developers"], "riskSource": [], "bountyEligibility": null}, "disclosurePlan": {"disclosureIntent": "No", "disclosureTimeline": null, "disclosureChannels": [], "embargoRequest": ""}, "malignActor": {"tactics": ["Execution"], "impact": []}, "securityIncident": {"threatActorIntent": null, "detection": ["Monitoring"]}}
|
ff23d9c1-c4cb-43ed-a4bc-ce4ffde0437a
|
Unknown
|
["Real-World Incidents"]
|
adsfasdf
|
2025-06-02T13:35:00.906359
|
{"Submitter Relationship": "Independent observer", "Submitter_Relationship_Other": "", "Incident Location(s)": "adswfcawdef", "Harm Narrative": "awfawsedf", "Disclosure Intent": "No", "Disclosure Timeline": null, "Disclosure Channels": [], "Disclosure_Channels_Other": "", "Embargo Request": "", "Report Types": ["Real-World Incidents"], "Reporter ID": "adsfasdf", "Session ID": "asdfadsf", "Flaw Timestamp Start": "2025-06-02", "Systems": ["Claude-3.7-Sonnet-Reasoning"], "Incident Description": "**Detailed Description:**\nasdfasdf\n\n**Undesirable Outputs/Effects/Impacts:**\nafasdwf\n\n**Reproduction Steps:**\nafdawdf\n\n**Systematic Evidence:**\nafawdsf", "Incident Description - Detailed": "asdfasdf", "Incident Description - Outputs": "afasdwf", "Incident Description - Reproduction": "afdawdf", "Incident Description - Systematic": "afawdsf", "Policy Violation": "dawsfawdsef", "Prevalence": "Rare", "Severity": "Medium", "Impacts": ["Privacy"], "Impacts_Other": "", "Impacted Stakeholder(s)": ["Vulnerable populations"], "Risk Source(s)": ["Implementation error"], "Context Info": "efawsef", "Report ID": "ff23d9c1-c4cb-43ed-a4bc-ce4ffde0437a", "Submission Timestamp": "2025-06-02T13:34:58.575598"}
|
{"@context": "https://schema.org", "@type": "AIFlawReport", "reportId": "ff23d9c1-c4cb-43ed-a4bc-ce4ffde0437a", "dateCreated": "2025-06-02T13:34:58.824580", "reportStatus": null, "reportTypes": ["Real-World Incidents"], "basicInformation": {"reporterId": "adsfasdf", "sessionId": "asdfadsf", "flawTimestampStart": "2025-06-02", "flawTimestampEnd": null, "systems": ["Claude-3.7-Sonnet-Reasoning"]}, "commonFields": {"contextInfo": "efawsef", "flawDescription": null, "policyViolation": "dawsfawdsef", "severity": "Medium", "prevalence": "Rare", "impacts": ["Privacy"], "impactedStakeholders": ["Vulnerable populations"], "riskSource": [], "bountyEligibility": null}, "disclosurePlan": {"disclosureIntent": "No", "disclosureTimeline": null, "disclosureChannels": [], "embargoRequest": ""}}
|
86a6fa1f-24d1-44f7-82ba-e5f944b7c4f4
|
Unknown
|
["Real-World Incidents"]
|
jehjefw
|
2025-06-02T13:49:29.528680
|
{"Submitter Relationship": "Independent observer", "Submitter_Relationship_Other": "", "Incident Location(s)": "awefawef", "Harm Narrative": "awedfcawedf", "Disclosure Intent": "No", "Disclosure Timeline": null, "Disclosure Channels": [], "Disclosure_Channels_Other": "", "Embargo Request": "", "Report Types": ["Real-World Incidents"], "Reporter ID": "jehjefw", "Session ID": "adsfdasf", "Flaw Timestamp Start": "2025-06-02", "Systems": ["Claude-3"], "Incident Description": "**Detailed Description:**\nasdfasdf\n\n**Undesirable Outputs/Effects/Impacts:**\nadfaf\n\n**Reproduction Steps:**\nafeawfedc\n\n**Systematic Evidence:**\nwadfawdf", "Incident Description - Detailed": "asdfasdf", "Incident Description - Outputs": "adfaf", "Incident Description - Reproduction": "afeawfedc", "Incident Description - Systematic": "wadfawdf", "Policy Violation": "awfawfe", "Prevalence": "Unknown", "Severity": "Low", "Impacts": ["Environmental"], "Impacts_Other": "", "Impacted Stakeholder(s)": ["Vulnerable populations"], "Risk Source(s)": ["Implementation error"], "Context Info": "waefawf", "Report ID": "86a6fa1f-24d1-44f7-82ba-e5f944b7c4f4", "Submission Timestamp": "2025-06-02T13:49:21.722271"}
|
{"@context": "https://schema.org", "@type": "AIFlawReport", "reportId": "86a6fa1f-24d1-44f7-82ba-e5f944b7c4f4", "dateCreated": "2025-06-02T13:49:21.856010", "reportStatus": null, "reportTypes": ["Real-World Incidents"], "basicInformation": {"reporterId": "jehjefw", "sessionId": "adsfdasf", "flawTimestampStart": "2025-06-02", "flawTimestampEnd": null, "systems": ["Claude-3"]}, "commonFields": {"contextInfo": "waefawf", "flawDescription": null, "policyViolation": "awfawfe", "severity": "Low", "prevalence": "Unknown", "impacts": ["Environmental"], "impactedStakeholders": ["Vulnerable populations"], "riskSource": [], "bountyEligibility": null}, "disclosurePlan": {"disclosureIntent": "No", "disclosureTimeline": null, "disclosureChannels": [], "embargoRequest": ""}}
|
8edde2bb-e8ef-4bfc-94d3-da1cf399201a
|
Unknown
|
["Real-World Incidents", "Malign Actor", "Security Incident Report"]
|
aswdfawdf
|
2025-06-02T17:26:15.790976
|
{"Submitter Relationship": "Affected stakeholder", "Submitter_Relationship_Other": "", "Incident Location(s)": "adsfasd", "Harm Narrative": "asdfdadsf", "Tactic Select": ["Discovery"], "Tactic_Select_Other": "", "Detection": ["Monitoring"], "Detection_Other": "", "Disclosure Intent": "No", "Disclosure Timeline": null, "Disclosure Channels": [], "Disclosure_Channels_Other": "", "Embargo Request": "", "Report Types": ["Real-World Incidents", "Malign Actor", "Security Incident Report"], "Reporter ID": "aswdfawdf", "Session ID": "asdfadsf", "Flaw Timestamp Start": "2025-06-02", "Systems": ["GPT-4", "GPT-4o"], "Incident Description": "**Detailed Description:**\nasdfadsf\n\n**Undesirable Outputs/Effects/Impacts:**\nasdfasdf\n\n**Reproduction Steps:**\nadfadsf\n\n**Systematic Evidence:**\nasdfasdf", "Incident Description - Detailed": "asdfadsf", "Incident Description - Outputs": "asdfasdf", "Incident Description - Reproduction": "adfadsf", "Incident Description - Systematic": "asdfasdf", "Policy Violation": "asdfasdf", "Prevalence": "Rare", "Severity": "Medium", "Impacts": ["Psychological"], "Impacts_Other": "", "Impacted Stakeholder(s)": ["General Public"], "Risk Source(s)": [], "Context Info": "asdfadsf", "Report ID": "8edde2bb-e8ef-4bfc-94d3-da1cf399201a", "Submission Timestamp": "2025-06-02T17:26:14.066919"}
|
{"@context": "https://schema.org", "@type": "AIFlawReport", "reportId": "8edde2bb-e8ef-4bfc-94d3-da1cf399201a", "dateCreated": "2025-06-02T17:26:14.215039", "reportStatus": null, "reportTypes": ["Real-World Incidents", "Malign Actor", "Security Incident Report"], "basicInformation": {"reporterId": "aswdfawdf", "sessionId": "asdfadsf", "flawTimestampStart": "2025-06-02", "flawTimestampEnd": null, "systems": ["GPT-4", "GPT-4o"]}, "commonFields": {"contextInfo": "asdfadsf", "flawDescription": null, "policyViolation": "asdfasdf", "severity": "Medium", "prevalence": "Rare", "impacts": ["Psychological"], "impactedStakeholders": ["General Public"], "riskSource": [], "bountyEligibility": null}, "disclosurePlan": {"disclosureIntent": "No", "disclosureTimeline": null, "disclosureChannels": [], "embargoRequest": ""}, "malignActor": {"tactics": ["Discovery"], "impact": []}, "securityIncident": {"threatActorIntent": null, "detection": ["Monitoring"]}}
|
89335e84-4d2d-417f-89a4-4edaa2a356b4
|
Unknown
|
["Real-World Incidents"]
|
asdfasdf
|
2025-06-02T17:40:05.851268
|
{"Submitter Relationship": "Affected stakeholder", "Submitter_Relationship_Other": "", "Incident Location(s)": "asdfasdf", "Harm Narrative": "adsfasdf", "Disclosure Intent": "No", "Disclosure Timeline": null, "Disclosure Channels": [], "Disclosure_Channels_Other": "", "Embargo Request": "", "Report Types": ["Real-World Incidents"], "Reporter ID": "asdfasdf", "Session ID": "asdfadsf", "Flaw Timestamp Start": "2025-06-02", "Systems": ["GPT-4", "Claude-3.7-Sonnet-Reasoning", "GPT-4o", "Claude-instant"], "Incident Description": "**Detailed Description:**\nfasdfasdf\n\n**Undesirable Outputs/Effects/Impacts:**\nsafasdf\n\n**Reproduction Steps:**\nasfdasdf\n\n**Systematic Evidence:**\nasdfasdf", "Incident Description - Detailed": "fasdfasdf", "Incident Description - Outputs": "safasdf", "Incident Description - Reproduction": "asfdasdf", "Incident Description - Systematic": "asdfasdf", "Policy Violation": "asdfasdf", "Prevalence": "Rare", "Severity": "Medium", "Impacts": ["Psychological"], "Impacts_Other": "", "Impacted Stakeholder(s)": ["Users"], "Risk Source(s)": ["Deployment issue"], "Context Info": "asdfasdf", "Report ID": "89335e84-4d2d-417f-89a4-4edaa2a356b4", "Submission Timestamp": "2025-06-02T17:40:04.221101"}
|
{"@context": "https://schema.org", "@type": "AIFlawReport", "reportId": "89335e84-4d2d-417f-89a4-4edaa2a356b4", "dateCreated": "2025-06-02T17:40:04.416782", "reportStatus": null, "reportTypes": ["Real-World Incidents"], "basicInformation": {"reporterId": "asdfasdf", "sessionId": "asdfadsf", "flawTimestampStart": "2025-06-02", "flawTimestampEnd": null, "systems": ["GPT-4", "Claude-3.7-Sonnet-Reasoning", "GPT-4o", "Claude-instant"]}, "commonFields": {"contextInfo": "asdfasdf", "flawDescription": null, "policyViolation": "asdfasdf", "severity": "Medium", "prevalence": "Rare", "impacts": ["Psychological"], "impactedStakeholders": ["Users"], "riskSource": [], "bountyEligibility": null}, "disclosurePlan": {"disclosureIntent": "No", "disclosureTimeline": null, "disclosureChannels": [], "embargoRequest": ""}}
|
26230142-d980-4d31-bcd4-30bff1d6d664
|
Unknown
|
["Real-World Incidents"]
|
jhlkjhlkj
|
2025-06-16T21:42:10.475736
|
{"Submitter Relationship": "Affected stakeholder", "Submitter_Relationship_Other": "", "Incident Location(s)": "asdfasdf", "Harm Narrative": "asdfasdf", "Disclosure Intent": "No", "Disclosure Timeline": null, "Disclosure Channels": [], "Disclosure_Channels_Other": "", "Embargo Request": "", "Report Types": ["Real-World Incidents"], "Reporter ID": "jhlkjhlkj", "Session ID": "fadsfasdf", "Flaw Timestamp Start": "2025-06-16", "Systems": ["Claude-3.7-Sonnet-Reasoning"], "Incident Description": "**Detailed Description:**\n fasdfasdf\n ", "Incident Description - Detailed": "fasdfasdf", "Policy Violation": "asdfasdf", "Prevalence": "Occasional", "Severity": "Medium", "Impacts": ["Reputational"], "Impacts_Other": "", "Impacted Stakeholder(s)": ["Organizations"], "Risk Source(s)": ["Data bias"], "Context Info": "asdfasdf", "Proof-of-Concept Exploit": "asdfasdf", "Report ID": "26230142-d980-4d31-bcd4-30bff1d6d664", "Submission Timestamp": "2025-06-16T21:42:08.765794"}
|
{"@context": "https://schema.org", "@type": "AIFlawReport", "reportId": "26230142-d980-4d31-bcd4-30bff1d6d664", "dateCreated": "2025-06-16T21:42:08.967783", "reportStatus": null, "reportTypes": ["Real-World Incidents"], "basicInformation": {"reporterId": "jhlkjhlkj", "sessionId": "fadsfasdf", "flawTimestampStart": "2025-06-16", "flawTimestampEnd": null, "systems": ["Claude-3.7-Sonnet-Reasoning"]}, "commonFields": {"contextInfo": "asdfasdf", "flawDescription": null, "policyViolation": "asdfasdf", "severity": "Medium", "prevalence": "Occasional", "impacts": ["Reputational"], "impactedStakeholders": ["Organizations"], "riskSource": [], "bountyEligibility": null}, "disclosurePlan": {"disclosureIntent": "No", "disclosureTimeline": null, "disclosureChannels": [], "embargoRequest": ""}}
|
54b5397c-aa34-4b77-94c9-9f2442501b98
|
Unknown
|
["Malign Actor", "Vulnerability Report"]
|
Anonymous
|
2025-06-20T18:35:48.606924
|
{"Attacker Resources": ["Direct query access \u2014 black-box \u2014 An attacker can query the system\u2014the degree of control can vary substantially (e.g., ability to control temperature, view logits, etc.)."], "Attacker Objectives": ["Abuse violation \u2014 An attacker uses an AI system for purposes that are harmful or otherwise unintended by the developer, often by evading guardrails."], "Objective Context": "A malicious actor could leverage this flaw for persuasive campaign communications.", "Disclosure Intent": "Yes", "Disclosure Timeline": "Medium-term (31-90 days)", "Disclosure Channels": ["Blog post"], "Disclosure_Channels_Other": "", "Embargo Request": "Delay until after election ", "Report Types": ["Malign Actor", "Vulnerability Report"], "Report ID": "54b5397c-aa34-4b77-94c9-9f2442501b98"}
|
{"@context": "https://schema.org", "@type": "AIFlawReport", "reportId": "54b5397c-aa34-4b77-94c9-9f2442501b98", "dateCreated": "2025-06-20T18:35:47.030420", "reportStatus": null, "reportTypes": ["Malign Actor", "Vulnerability Report"], "basicInformation": {"reporterId": null, "sessionId": null, "flawTimestampStart": null, "flawTimestampEnd": null, "systems": []}, "commonFields": {"contextInfo": null, "flawDescription": null, "policyViolation": null, "severity": null, "prevalence": null, "impacts": [], "impactedStakeholders": [], "riskSource": [], "bountyEligibility": null}, "disclosurePlan": {"disclosureIntent": "Yes", "disclosureTimeline": "Medium-term (31-90 days)", "disclosureChannels": ["Blog post"], "embargoRequest": "Delay until after election "}, "malignActor": {"tactics": [], "impact": []}, "vulnerability": {"proofOfConcept": null}}
|
1aa9ec82-0a0a-4bf3-b498-fc2338921f04
|
Unknown
|
["Hazard Report"]
|
Anonymous
|
2025-07-19T17:34:28.663671
|
{"Statistical Argument with Examples": null, "Disclosure Intent": "Yes", "Disclosure Timeline": "Immediate (0 days)", "Disclosure Channels": [], "Disclosure_Channels_Other": "", "Embargo Request": null, "Report Types": ["Hazard Report"], "Report ID": "1aa9ec82-0a0a-4bf3-b498-fc2338921f04"}
|
{"@context": "https://schema.org", "@type": "AIFlawReport", "reportId": "1aa9ec82-0a0a-4bf3-b498-fc2338921f04", "dateCreated": "2025-07-19T17:34:27.245698", "reportStatus": null, "reportTypes": ["Hazard Report"], "basicInformation": {"reporterId": null, "sessionId": null, "flawTimestampStart": null, "flawTimestampEnd": null, "systems": []}, "commonFields": {"contextInfo": null, "flawDescription": null, "policyViolation": null, "severity": null, "prevalence": null, "impacts": [], "impactedStakeholders": [], "riskSource": [], "bountyEligibility": null}, "disclosurePlan": {"disclosureIntent": "Yes", "disclosureTimeline": "Immediate (0 days)", "disclosureChannels": [], "embargoRequest": null}, "hazard": {"examples": null, "replicationPacket": null, "statisticalArgument": null}}
|
27284a68-fabd-4fd8-8ebe-7d9ce0f2be30
|
Unknown
|
["Real-World Incidents", "Malign Actor", "Security Incident Report"]
|
anonymous-1518497c
|
2025-07-19T17:35:08.630261
|
{"Submitter Relationship": "Affected stakeholder", "Submitter_Relationship_Other": "", "Incident Location(s)": null, "Harm Narrative": null, "Attacker Resources": [], "Attacker Objectives": [], "Objective Context": "", "Detection": [], "Detection_Other": "", "Disclosure Intent": "Yes", "Disclosure Timeline": "Immediate (0 days)", "Disclosure Channels": [], "Disclosure_Channels_Other": "", "Embargo Request": null, "Report Types": ["Real-World Incidents", "Malign Actor", "Security Incident Report"], "Reporter ID": "anonymous-1518497c", "Session ID": null, "Flaw Timestamp Start": "2025-07-18", "Systems": ["GPT-4o"], "Incident Description": "**Detailed Description:**\n Test\n ", "Incident Description - Detailed": "Test", "Policy Violation": "Test", "Prevalence": "Unknown", "Severity": "Low", "Impacts": ["Economic/property"], "Impacts_Other": "", "CSAM Related": null, "Specific Harm Types": [], "Impacted Stakeholder(s)": ["Administrators"], "Risk Source(s)": {"Responsible Factors": [], "Responsible Factors Subcategories": {}, "Responsible Factors Context": ""}, "Context Info": null, "Proof-of-Concept Exploit": null, "Report ID": "27284a68-fabd-4fd8-8ebe-7d9ce0f2be30", "Submission Timestamp": "2025-07-19T17:35:06.919952"}
|
{"@context": "https://schema.org", "@type": "AIFlawReport", "reportId": "27284a68-fabd-4fd8-8ebe-7d9ce0f2be30", "dateCreated": "2025-07-19T17:35:07.040137", "reportStatus": null, "reportTypes": ["Real-World Incidents", "Malign Actor", "Security Incident Report"], "basicInformation": {"reporterId": "anonymous-1518497c", "sessionId": null, "flawTimestampStart": "2025-07-18", "flawTimestampEnd": null, "systems": ["GPT-4o"]}, "commonFields": {"contextInfo": null, "flawDescription": null, "policyViolation": "Test", "severity": "Low", "prevalence": "Unknown", "impacts": ["Economic/property"], "impactedStakeholders": ["Administrators"], "riskSource": [], "bountyEligibility": null}, "disclosurePlan": {"disclosureIntent": "Yes", "disclosureTimeline": "Immediate (0 days)", "disclosureChannels": [], "embargoRequest": null}, "malignActor": {"tactics": [], "impact": []}, "securityIncident": {"threatActorIntent": null, "detection": []}}
|
ccb8a081-593f-4073-85a4-6527419a94c9
|
Unknown
|
["Real-World Incidents", "Malign Actor", "Security Incident Report"]
|
anonymous-63ddae10
|
2025-07-21T14:58:27.015151
|
{"Submitter Relationship": "Affected stakeholder", "Submitter_Relationship_Other": "", "Incident Location(s)": null, "Harm Narrative": null, "Attacker Resources": [], "Attacker Objectives": [], "Objective Context": "", "Detection": [], "Detection_Other": "", "Disclosure Intent": "Yes", "Disclosure Timeline": "Immediate (0 days)", "Disclosure Channels": [], "Disclosure_Channels_Other": "", "Embargo Request": null, "Report Types": ["Real-World Incidents", "Malign Actor", "Security Incident Report"], "Reporter ID": "anonymous-63ddae10", "Session ID": null, "Flaw Timestamp Start": "2025-07-16", "Systems": ["GPT-4", "o3-mini", "Claude-3.7-Sonnet"], "Incident Description": "**Detailed Description:**\n test\n ", "Incident Description - Detailed": "test", "Policy Violation": "test", "Prevalence": "Unknown", "Severity": "Low", "Impacts": ["Physical"], "Impacts_Other": "", "CSAM Related": null, "Specific Harm Types": ["Property damage - Action(s) that lead directly or indirectly to the damage or destruction of tangible property eg. buildings, possessions, vehicles, robots"], "Impacted Stakeholder(s)": ["Administrators"], "Risk Source(s)": {"Responsible Factors": [], "Responsible Factors Subcategories": {}, "Responsible Factors Context": ""}, "Context Info": null, "Proof-of-Concept Exploit": null, "Report ID": "ccb8a081-593f-4073-85a4-6527419a94c9", "Submission Timestamp": "2025-07-21T14:58:24.889766"}
|
{"@context": "https://schema.org", "@type": "AIFlawReport", "reportId": "ccb8a081-593f-4073-85a4-6527419a94c9", "dateCreated": "2025-07-21T14:58:25.217109", "reportStatus": null, "reportTypes": ["Real-World Incidents", "Malign Actor", "Security Incident Report"], "basicInformation": {"reporterId": "anonymous-63ddae10", "sessionId": null, "flawTimestampStart": "2025-07-16", "flawTimestampEnd": null, "systems": ["GPT-4", "o3-mini", "Claude-3.7-Sonnet"]}, "commonFields": {"contextInfo": null, "flawDescription": null, "policyViolation": "test", "severity": "Low", "prevalence": "Unknown", "impacts": ["Physical"], "impactedStakeholders": ["Administrators"], "riskSource": [], "bountyEligibility": null}, "disclosurePlan": {"disclosureIntent": "Yes", "disclosureTimeline": "Immediate (0 days)", "disclosureChannels": [], "embargoRequest": null}, "malignActor": {"tactics": [], "impact": []}, "securityIncident": {"threatActorIntent": null, "detection": []}}
|
f38f5cea-3686-4e36-a347-42ab11f9944b
|
Unknown
|
["Malign Actor", "Vulnerability Report"]
|
anonymous-8ffd265a
|
2025-08-04T08:58:29.276877
|
{"Attacker Resources": ["Model/system supply chain control \u2014 An attacker can modify the AI model itself, such as via public fine-tuning"], "Attacker Objectives": ["Privacy compromise \u2014 An attacker gains access to sensitive and confidential information, including information about the AI system (e.g., architecture or weights) or sensitive information that the model accesses (e.g., training data, external knowledge databases."], "Objective Context": null, "Disclosure Intent": "Yes", "Disclosure Timeline": "Medium-term (31-90 days)", "Disclosure Channels": ["Blog post", "Media outlet"], "Disclosure_Channels_Other": "", "Embargo Request": null, "Report Types": ["Malign Actor", "Vulnerability Report"], "Reporter ID": "anonymous-8ffd265a", "Session ID": null, "Flaw Timestamp Start": "2025-08-20", "Systems": ["GPT-4o"], "Flaw Description": "**Detailed Description:**\n \n ", "Flaw Description - Detailed": null, "Policy Violation": "all of them", "Prevalence": "Widespread", "Severity": "Critical", "Impacts": ["Psychological", "Physical"], "Impacts_Other": "", "CSAM Related": null, "Specific Harm Types": ["Anxiety/depression - Mental health decline due to addiction, negative social interactions such as humiliation and shaming and traumatic distressing events such as online violence or rape", "Radicalisation - Adoption of extreme political, social, or religious ideals and aspirations due to the nature or misuse of an algorithmic system, potentially resulting in abuse, violence, or terrorism"], "Impacted Stakeholder(s)": ["Vulnerable populations"], "Risk Source(s)": {"Responsible Factors": ["Feedback", "Supply chain weaknesses (e.g., software libraries and hardware)"], "Responsible Factors Subcategories": {"Feedback": ["Poisoning", "Misspecification (no threat actor)"]}, "Responsible Factors Context": null}, "Context Info": null, "Proof-of-Concept Exploit": null, "Report ID": "f38f5cea-3686-4e36-a347-42ab11f9944b", "Submission Timestamp": "2025-08-04T08:58:25.237831"}
|
{"@context": "https://schema.org", "@type": "AIFlawReport", "reportId": "f38f5cea-3686-4e36-a347-42ab11f9944b", "dateCreated": "2025-08-04T08:58:25.893246", "reportStatus": null, "reportTypes": ["Malign Actor", "Vulnerability Report"], "basicInformation": {"reporterId": "anonymous-8ffd265a", "sessionId": null, "flawTimestampStart": "2025-08-20", "flawTimestampEnd": null, "systems": ["GPT-4o"]}, "commonFields": {"contextInfo": null, "flawDescription": "**Detailed Description:**\n \n ", "policyViolation": "all of them", "severity": "Critical", "prevalence": "Widespread", "impacts": ["Psychological", "Physical"], "impactedStakeholders": ["Vulnerable populations"], "riskSource": [], "bountyEligibility": null}, "disclosurePlan": {"disclosureIntent": "Yes", "disclosureTimeline": "Medium-term (31-90 days)", "disclosureChannels": ["Blog post", "Media outlet"], "embargoRequest": null}, "malignActor": {"tactics": [], "impact": []}, "vulnerability": {"proofOfConcept": null}}
|
ea81fb6e-b283-4cad-b414-3e82ef29d03b
|
Unknown
|
["Malign Actor", "Vulnerability Report"]
|
@LChoshen
|
2025-08-18T12:44:54.744810
|
{"Attacker Resources": ["Training data/feedback control \u2014 An attacker can modify training data/feedback and/or insert a subset of examples.", "Model/system supply chain control \u2014 An attacker can modify the AI model itself, such as via public fine-tuning", "Direct query access \u2014 black-box \u2014 An attacker can query the system\u2014the degree of control can vary substantially (e.g., ability to control temperature, view logits, etc.)."], "Attacker Objectives": ["Integrity violation \u2014 An attacker causes AI systems to perform tasks inadequately or behave undesirably.", "Abuse violation \u2014 An attacker uses an AI system for purposes that are harmful or otherwise unintended by the developer, often by evading guardrails.", "Availability breakdown \u2014 An attacker disrupts the ability of users to obtain access to AI systems or their functionality."], "Objective Context": null, "Disclosure Intent": "Already Public Knowledge", "Disclosure Timeline": null, "Disclosure Channels": [], "Disclosure_Channels_Other": "", "Embargo Request": "", "Report Types": ["Malign Actor", "Vulnerability Report"], "Reporter ID": "@LChoshen", "Session ID": null, "Flaw Timestamp Start": "2025-06-25", "Systems": ["GPT-4.5-Preview", "o3-mini", "o1", "GPT-4o", "GPT-4", "GPT-3.5-Turbo", "Claude-3.7-Sonnet-Reasoning"], "Flaw Description": "**Detailed Description:**\n Malicious feedback can force the model to say things not previously in the model training, to prefer them and to make the model learn them and use them in non-malicious cases. It is hard to say to which extent this affect each provider as their practices are not disclosed.\nWe found that one can sway models to output bugs, unwanted code, fake information, etc. it is done by making the model repeat the information we want it to (e.g. \"pick one of the following sentences and say them randomly [ans a, ans b] then a benign instruction\")\n\n ", "Flaw Description - Detailed": "Malicious feedback can force the model to say things not previously in the model training, to prefer them and to make the model learn them and use them in non-malicious cases. It is hard to say to which extent this affect each provider as their practices are not disclosed.\nWe found that one can sway models to output bugs, unwanted code, fake information, etc. it is done by making the model repeat the information we want it to (e.g. \"pick one of the following sentences and say them randomly [ans a, ans b] then a benign instruction\")\n", "Policy Violation": "The AI would break whatever you want. It will help my brand, give false information, lead people to press on my malicious link, import wrong libraries (that I can create to wrap real ones + a security break) etc.", "Prevalence": "Unknown", "Severity": "Critical", "Impacts": ["Autonomy", "Physical", "Psychological", "Reputational", "Financial and Business", "Human Rights and Civil Liberties", "Societal and Cultural", "Political and Economic", "Environmental", "Sexualization"], "Impacts_Other": "", "CSAM Related": "No", "Specific Harm Types": [], "Impacted Stakeholder(s)": ["Users"], "Risk Source(s)": {"Responsible Factors": [], "Responsible Factors Subcategories": {}, "Responsible Factors Context": ""}, "Context Info": "https://arxiv.org/abs/2507.02850", "Proof-of-Concept Exploit": null, "Report ID": "ea81fb6e-b283-4cad-b414-3e82ef29d03b", "Submission Timestamp": "2025-08-18T12:44:52.920304"}
|
{"@context": "https://schema.org", "@type": "AIFlawReport", "reportId": "ea81fb6e-b283-4cad-b414-3e82ef29d03b", "dateCreated": "2025-08-18T12:44:53.118987", "reportStatus": null, "reportTypes": ["Malign Actor", "Vulnerability Report"], "basicInformation": {"reporterId": "@LChoshen", "sessionId": null, "flawTimestampStart": "2025-06-25", "flawTimestampEnd": null, "systems": ["GPT-4.5-Preview", "o3-mini", "o1", "GPT-4o", "GPT-4", "GPT-3.5-Turbo", "Claude-3.7-Sonnet-Reasoning"]}, "commonFields": {"contextInfo": "https://arxiv.org/abs/2507.02850", "flawDescription": "**Detailed Description:**\n Malicious feedback can force the model to say things not previously in the model training, to prefer them and to make the model learn them and use them in non-malicious cases. It is hard to say to which extent this affect each provider as their practices are not disclosed.\nWe found that one can sway models to output bugs, unwanted code, fake information, etc. it is done by making the model repeat the information we want it to (e.g. \"pick one of the following sentences and say them randomly [ans a, ans b] then a benign instruction\")\n\n ", "policyViolation": "The AI would break whatever you want. It will help my brand, give false information, lead people to press on my malicious link, import wrong libraries (that I can create to wrap real ones + a security break) etc.", "severity": "Critical", "prevalence": "Unknown", "impacts": ["Autonomy", "Physical", "Psychological", "Reputational", "Financial and Business", "Human Rights and Civil Liberties", "Societal and Cultural", "Political and Economic", "Environmental", "Sexualization"], "impactedStakeholders": ["Users"], "riskSource": [], "bountyEligibility": null}, "disclosurePlan": {"disclosureIntent": "Already Public Knowledge", "disclosureTimeline": null, "disclosureChannels": [], "embargoRequest": ""}, "malignActor": {"tactics": [], "impact": []}, "vulnerability": {"proofOfConcept": null}}
|
097e8ac1-2efd-4e11-8de4-34db667ad5ca
|
Unknown
|
["Real-World Incidents"]
|
Anonymous
|
2025-08-19T15:47:09.485188
|
{"Submitter Relationship": "Affected stakeholder", "Submitter_Relationship_Other": "", "Incident Location(s)": "fadsfasdf", "Harm Narrative": "asdfasdf", "Disclosure Intent": "No", "Disclosure Timeline": null, "Disclosure Channels": [], "Disclosure_Channels_Other": "", "Embargo Request": "", "Report Types": ["Real-World Incidents"], "Report ID": "097e8ac1-2efd-4e11-8de4-34db667ad5ca"}
|
"{\n \"@context\": [\n \"https://schema.org/\",\n {\n \"aifr\": \"https://aiflawreports.org/schema/\",\n \"aiSystem\": \"aifr:aiSystem\",\n \"severity\": \"aifr:severity\",\n \"prevalence\": \"aifr:prevalence\",\n \"impacts\": \"aifr:impacts\",\n \"reportType\": \"aifr:reportType\"\n }\n ],\n \"@type\": \"aifr:AIFlawReport\",\n \"@id\": \"https://aiflawreports.org/reports/097e8ac1-2efd-4e11-8de4-34db667ad5ca\",\n \"name\": \"AI Flaw Report: Unknown System\",\n \"description\": \"No description provided\",\n \"aiSystem\": [\n {\n \"@type\": \"schema:SoftwareApplication\",\n \"@id\": \"https://aiflawreports.org/reports/097e8ac1-2efd-4e11-8de4-34db667ad5ca/unknown-system\",\n \"name\": \"Unknown System\",\n \"description\": \"No specific system identified\"\n }\n ],\n \"severity\": \"Unknown\",\n \"prevalence\": \"Unknown\",\n \"impacts\": [],\n \"reportType\": [\n \"Real-World Incidents\"\n ],\n \"dateCreated\": \"2025-08-19T19:47:05.255809+00:00\",\n \"identifier\": \"097e8ac1-2efd-4e11-8de4-34db667ad5ca\",\n \"aifr:incident\": {\n \"@type\": \"aifr:RealWorldIncident\",\n \"description\": null,\n \"location\": \"fadsfasdf\",\n \"aifr:harmNarrative\": \"asdfasdf\"\n },\n \"aifr:disclosure\": {\n \"@type\": \"aifr:DisclosurePlan\",\n \"aifr:intent\": \"No\"\n }\n}"
|
df785d32-58d9-470e-bbd1-1dd0e5b490b8
|
Unknown
|
["Real-World Incidents"]
|
Anonymous
|
2025-08-19T15:48:40.422982
|
{"Submitter Relationship": "Affected stakeholder", "Submitter_Relationship_Other": "", "Incident Location(s)": "dfasdf", "Harm Narrative": "asdfasdf", "Disclosure Intent": "No", "Disclosure Timeline": null, "Disclosure Channels": [], "Disclosure_Channels_Other": "", "Embargo Request": "", "Report Types": ["Real-World Incidents"], "Report ID": "df785d32-58d9-470e-bbd1-1dd0e5b490b8"}
|
{"@context": "https://schema.org", "@type": "AIFlawReport", "reportId": "df785d32-58d9-470e-bbd1-1dd0e5b490b8", "dateCreated": "2025-08-19T15:48:38.290354", "reportStatus": null, "reportTypes": ["Real-World Incidents"], "basicInformation": {"reporterId": null, "sessionId": null, "flawTimestampStart": null, "flawTimestampEnd": null, "systems": []}, "commonFields": {"contextInfo": null, "flawDescription": null, "policyViolation": null, "severity": null, "prevalence": null, "impacts": [], "impactedStakeholders": [], "riskSource": [], "bountyEligibility": null}, "disclosurePlan": {"disclosureIntent": "No", "disclosureTimeline": null, "disclosureChannels": [], "embargoRequest": ""}}
|
cf984817-e13a-4d74-bd2b-2ff8570a774b
|
Unknown
|
["Real-World Incidents"]
|
Anonymous
|
2025-08-19T15:49:49.763793
|
{"Submitter Relationship": "Affected stakeholder", "Submitter_Relationship_Other": "", "Incident Location(s)": "adsfasdf", "Harm Narrative": "asdfasdf", "Disclosure Intent": "No", "Disclosure Timeline": null, "Disclosure Channels": [], "Disclosure_Channels_Other": "", "Embargo Request": "", "Report Types": ["Real-World Incidents"], "Report ID": "cf984817-e13a-4d74-bd2b-2ff8570a774b"}
|
{"@context": "https://schema.org", "@type": "AIFlawReport", "reportId": "cf984817-e13a-4d74-bd2b-2ff8570a774b", "dateCreated": "2025-08-19T15:49:47.352245", "reportStatus": null, "reportTypes": ["Real-World Incidents"], "basicInformation": {"reporterId": null, "sessionId": null, "flawTimestampStart": null, "flawTimestampEnd": null, "systems": []}, "commonFields": {"contextInfo": null, "flawDescription": null, "policyViolation": null, "severity": null, "prevalence": null, "impacts": [], "impactedStakeholders": [], "riskSource": [], "bountyEligibility": null}, "disclosurePlan": {"disclosureIntent": "No", "disclosureTimeline": null, "disclosureChannels": [], "embargoRequest": ""}}
|
54c4347b-24e7-41d5-8c9b-c5c23b62dc41
|
Unknown
|
["Real-World Incidents"]
|
asdfasdf
|
2025-08-19T16:09:22.656054
|
{"Submitter Relationship": "Affected stakeholder", "Submitter_Relationship_Other": "", "Incident Location(s)": "fdsfsdf", "Harm Narrative": "asdfasdf", "Disclosure Intent": "No", "Disclosure Timeline": null, "Disclosure Channels": [], "Disclosure_Channels_Other": "", "Embargo Request": "", "Report Types": ["Real-World Incidents"], "Reporter ID": "asdfasdf", "Session ID": "asdfasdf", "Flaw Timestamp Start": "2025-08-19", "Systems": ["o3-mini"], "Incident Description": "**Detailed Description:**\n fdsdfsd\n ", "Incident Description - Detailed": "fdsdfsd", "Potential Policy Violation": "sdfsdfa", "Prevalence": "Occasional", "Severity": "Significant", "Impacts": ["Public interest/critical infrastructure"], "Impacts_Other": "", "CSAM Related": null, "Specific Harm Types": [], "Impacted Stakeholder(s)": ["General Public"], "Risk Source(s)": {"Responsible Factors": ["Training data"], "Responsible Factors Subcategories": {"Training data": ["Data poisoning"]}, "Responsible Factors Context": "asdfasd"}, "Context Info": "asdfasdf", "Proof-of-Concept Exploit": "asdfasdf", "Report ID": "54c4347b-24e7-41d5-8c9b-c5c23b62dc41", "Submission Timestamp": "2025-08-19T15:53:26.053576"}
|
"{\n \"@context\": \"https://schema.org/\",\n \"@type\": \"Report\",\n \"@id\": \"https://aiflawreports.org/reports/54c4347b-24e7-41d5-8c9b-c5c23b62dc41\",\n \"name\": \"AI Flaw Report\",\n \"description\": \"\",\n \"dateCreated\": \"2025-08-19T20:09:20.806348+00:00\",\n \"aifr:processingError\": \"\\\"ProcessedAIFlawReport\\\" object has no field \\\"raw_data\\\"\"\n}"
|
f30aa923-5a9c-4223-bbc2-9387365a4de2
|
Unknown
|
["Real-World Incidents"]
|
asdfasdf
|
2025-08-19T16:10:04.973130
|
{"Submitter Relationship": "Affected stakeholder", "Submitter_Relationship_Other": "", "Incident Location(s)": "adfasdf", "Harm Narrative": "adsfadf", "Disclosure Intent": "No", "Disclosure Timeline": null, "Disclosure Channels": [], "Disclosure_Channels_Other": "", "Embargo Request": "", "Report Types": ["Real-World Incidents"], "Reporter ID": "asdfasdf", "Session ID": "asdfasdf", "Flaw Timestamp Start": "2025-08-19", "Systems": ["GPT-3.5-Turbo"], "Incident Description": "**Detailed Description:**\n asdgasdg\n ", "Incident Description - Detailed": "asdgasdg", "Potential Policy Violation": "asdgasdg", "Prevalence": "Rare", "Severity": "Medium", "Impacts": ["Environmental"], "Impacts_Other": "", "CSAM Related": null, "Specific Harm Types": ["Biodiversity loss - Over-expansion of technology infrastructure, or inadequate alignment of technology with sustainable practices, leading to deforestation, habitat destruction, and fragmentation and loss of biodiversity"], "Impacted Stakeholder(s)": ["Developers"], "Risk Source(s)": {"Responsible Factors": ["Memory"], "Responsible Factors Subcategories": {"Memory": ["Indirect prompt injection (memory poisoning)"]}, "Responsible Factors Context": "adgasdg"}, "Context Info": "adgasdg", "Proof-of-Concept Exploit": "agsdgasdg", "Report ID": "f30aa923-5a9c-4223-bbc2-9387365a4de2", "Submission Timestamp": "2025-08-19T16:09:56.998965"}
|
"{\n \"@context\": \"https://schema.org/\",\n \"@type\": \"Report\",\n \"@id\": \"https://aiflawreports.org/reports/f30aa923-5a9c-4223-bbc2-9387365a4de2\",\n \"name\": \"AI Flaw Report\",\n \"description\": \"\",\n \"dateCreated\": \"2025-08-19T20:10:02.833948+00:00\",\n \"aifr:processingError\": \"\\\"ProcessedAIFlawReport\\\" object has no field \\\"raw_data\\\"\"\n}"
|
f4c0ae9d-bcb6-4752-a762-27e07c321821
|
Unknown
|
["Hazard Report"]
|
asdgasdg
|
2025-08-19T16:12:42.833994
|
{"Statistical Argument with Examples": "adsfasdf", "Disclosure Intent": "No", "Disclosure Timeline": null, "Disclosure Channels": [], "Disclosure_Channels_Other": "", "Embargo Request": "", "Report Types": ["Hazard Report"], "Reporter ID": "asdgasdg", "Session ID": "asdgasd", "Flaw Timestamp Start": "2025-08-12", "Systems": ["o3-mini"], "Flaw Description": "**Detailed Description:**\n fasdfasdf\n ", "Flaw Description - Detailed": "fasdfasdf", "Policy Violation": "adsfasdf", "Prevalence": "Occasional", "Severity": "Low", "Impacts": ["Autonomy"], "Impacts_Other": "", "CSAM Related": null, "Specific Harm Types": ["Autonomy/agency loss - Loss of an individual, group or organisation's ability to make informed decisions or pursue goals"], "Impacted Stakeholder(s)": ["Users"], "Risk Source(s)": {"Responsible Factors": ["Training data"], "Responsible Factors Subcategories": {"Training data": ["Sensitive information (e.g., PII)"]}, "Responsible Factors Context": "asdgasd"}, "Context Info": "adfasdf", "Proof-of-Concept Exploit": "asdfasdf", "Report ID": "f4c0ae9d-bcb6-4752-a762-27e07c321821", "Submission Timestamp": "2025-08-19T16:12:33.655075"}
|
"{\n \"@context\": [\n \"https://schema.org/\",\n {\n \"aifr\": \"https://aiflawreports.org/schema/\",\n \"aiSystem\": \"aifr:aiSystem\",\n \"severity\": \"aifr:severity\",\n \"prevalence\": \"aifr:prevalence\",\n \"impacts\": \"aifr:impacts\",\n \"reportType\": \"aifr:reportType\",\n \"riskSource\": \"aifr:riskSource\",\n \"contextInfo\": \"aifr:contextInfo\"\n }\n ],\n \"@type\": \"aifr:AIFlawReport\",\n \"@id\": \"https://aiflawreports.org/reports/f4c0ae9d-bcb6-4752-a762-27e07c321821\",\n \"name\": \"AI Flaw Report: o3-mini\",\n \"description\": \"**Detailed Description:**\\n fasdfasdf\\n \",\n \"aiSystem\": [\n {\n \"@type\": \"schema:SoftwareApplication\",\n \"@id\": \"https://aiflawreports.org/systems/o3-mini\",\n \"name\": \"o3-mini\",\n \"version\": \"\"\n }\n ],\n \"severity\": \"Low\",\n \"prevalence\": \"Occasional\",\n \"impacts\": [\n \"Autonomy\"\n ],\n \"reportType\": [\n \"Hazard Report\"\n ],\n \"dateCreated\": \"2025-08-19T20:12:40.943427+00:00\",\n \"identifier\": \"f4c0ae9d-bcb6-4752-a762-27e07c321821\",\n \"aifr:policyViolation\": \"adsfasdf\",\n \"author\": {\n \"@type\": \"schema:Person\",\n \"identifier\": \"asdgasdg\"\n },\n \"aifr:sessionId\": \"asdgasd\",\n \"aifr:flawTimestamp\": \"2025-08-12\",\n \"aifr:impactedStakeholders\": [\n \"Users\"\n ],\n \"aifr:specificHarmTypes\": [\n \"Autonomy/agency loss - Loss of an individual, group or organisation's ability to make informed decisions or pursue goals\"\n ],\n \"aifr:hazard\": {\n \"@type\": \"aifr:Hazard\",\n \"aifr:statisticalArgument\": \"adsfasdf\"\n },\n \"aifr:disclosure\": {\n \"@type\": \"aifr:DisclosurePlan\",\n \"aifr:intent\": \"No\"\n },\n \"aifr:raw\": {\n \"Statistical Argument with Examples\": \"adsfasdf\",\n \"Disclosure Intent\": \"No\",\n \"Disclosure Timeline\": null,\n \"Disclosure Channels\": [],\n \"Disclosure_Channels_Other\": \"\",\n \"Embargo Request\": \"\",\n \"Report Types\": [\n \"Hazard Report\"\n ],\n \"Reporter ID\": \"asdgasdg\",\n \"Session ID\": \"asdgasd\",\n \"Flaw Timestamp Start\": \"2025-08-12\",\n \"Systems\": [\n \"o3-mini\"\n ],\n \"Flaw Description\": \"**Detailed Description:**\\n fasdfasdf\\n \",\n \"Flaw Description - Detailed\": \"fasdfasdf\",\n \"Policy Violation\": \"adsfasdf\",\n \"Prevalence\": \"Occasional\",\n \"Severity\": \"Low\",\n \"Impacts\": [\n \"Autonomy\"\n ],\n \"Impacts_Other\": \"\",\n \"CSAM Related\": null,\n \"Specific Harm Types\": [\n \"Autonomy/agency loss - Loss of an individual, group or organisation's ability to make informed decisions or pursue goals\"\n ],\n \"Impacted Stakeholder(s)\": [\n \"Users\"\n ],\n \"Risk Source(s)\": {\n \"Responsible Factors\": [\n \"Training data\"\n ],\n \"Responsible Factors Subcategories\": {\n \"Training data\": [\n \"Sensitive information (e.g., PII)\"\n ]\n },\n \"Responsible Factors Context\": \"asdgasd\"\n },\n \"Context Info\": \"adfasdf\",\n \"Proof-of-Concept Exploit\": \"asdfasdf\",\n \"Report ID\": \"f4c0ae9d-bcb6-4752-a762-27e07c321821\",\n \"Submission Timestamp\": \"2025-08-19T16:12:33.655075\"\n }\n}"
|
ef6fa01c-fccd-4430-b81f-d8953d3f70f0
|
Unknown
|
["Real-World Incidents", "Malign Actor", "Security Incident Report"]
|
asdfasdf
|
2025-08-19T16:16:26.380829
|
{"Submitter Relationship": "Affected stakeholder", "Submitter_Relationship_Other": "", "Incident Location(s)": "adsgasd", "Harm Narrative": "asdgadsf", "Attacker Resources": ["Training data/feedback control \u2014 An attacker can modify training data/feedback and/or insert a subset of examples."], "Attacker Objectives": ["Availability breakdown \u2014 An attacker disrupts the ability of users to obtain access to AI systems or their functionality."], "Objective Context": "asdfasdf", "Detection": ["User observation"], "Detection_Other": "", "Disclosure Intent": "No", "Disclosure Timeline": null, "Disclosure Channels": [], "Disclosure_Channels_Other": "", "Embargo Request": "", "Report Types": ["Real-World Incidents", "Malign Actor", "Security Incident Report"], "Reporter ID": "asdfasdf", "Session ID": "asdfasdf", "Flaw Timestamp Start": "2025-08-19", "Systems": ["GPT-3.5-Turbo"], "Incident Description": "**Detailed Description:**\n adsfasdf\n ", "Incident Description - Detailed": "adsfasdf", "Potential Policy Violation": "asdgasdg", "Prevalence": "Rare", "Severity": "Medium", "Impacts": ["Environmental"], "Impacts_Other": "", "CSAM Related": null, "Specific Harm Types": ["Excessive landfill - Excessive disposal of electrical or electronic equipment leading to ecological/biodiversity damage, and disrupting the livelihoods and eroding the rights of local communities"], "Impacted Stakeholder(s)": ["Users"], "Risk Source(s)": {"Responsible Factors": ["Tool outputs/external inputs"], "Responsible Factors Subcategories": {"Tool outputs/external inputs": ["Resource overload/accessibility issues"]}, "Responsible Factors Context": "fasdfasdf"}, "Context Info": "adfasdf", "Proof-of-Concept Exploit": "asdfasdf", "Report ID": "ef6fa01c-fccd-4430-b81f-d8953d3f70f0", "Submission Timestamp": "2025-08-19T16:16:13.583287"}
|
"{\n \"@context\": [\n \"https://schema.org/\",\n {\n \"aifr\": \"https://aiflawreports.org/schema/\",\n \"aiSystem\": \"aifr:aiSystem\",\n \"severity\": \"aifr:severity\",\n \"prevalence\": \"aifr:prevalence\",\n \"impacts\": \"aifr:impacts\",\n \"reportType\": \"aifr:reportType\",\n \"riskSource\": \"aifr:riskSource\",\n \"contextInfo\": \"aifr:contextInfo\"\n }\n ],\n \"@type\": \"aifr:AIFlawReport\",\n \"@id\": \"https://aiflawreports.org/reports/ef6fa01c-fccd-4430-b81f-d8953d3f70f0\",\n \"name\": \"AI Flaw Report: GPT-3.5-Turbo\",\n \"description\": \"adsfasdf\\n \",\n \"aiSystem\": [\n {\n \"@type\": \"schema:SoftwareApplication\",\n \"@id\": \"https://aiflawreports.org/systems/GPT-3.5-Turbo\",\n \"name\": \"GPT-3.5-Turbo\",\n \"version\": \"\"\n }\n ],\n \"severity\": \"Medium\",\n \"prevalence\": \"Rare\",\n \"impacts\": [\n \"Environmental\"\n ],\n \"reportType\": [\n \"Real-World Incidents\",\n \"Malign Actor\",\n \"Security Incident Report\"\n ],\n \"dateCreated\": \"2025-08-19T20:16:23.971140+00:00\",\n \"identifier\": \"ef6fa01c-fccd-4430-b81f-d8953d3f70f0\",\n \"aifr:policyViolation\": \"asdgasdg\",\n \"author\": {\n \"@type\": \"schema:Person\",\n \"identifier\": \"asdfasdf\"\n },\n \"aifr:sessionId\": \"asdfasdf\",\n \"aifr:flawTimestamp\": \"2025-08-19\",\n \"aifr:impactedStakeholders\": [\n \"Users\"\n ],\n \"aifr:specificHarmTypes\": [\n \"Excessive landfill - Excessive disposal of electrical or electronic equipment leading to ecological/biodiversity damage, and disrupting the livelihoods and eroding the rights of local communities\"\n ],\n \"aifr:incident\": {\n \"@type\": \"aifr:RealWorldIncident\",\n \"description\": \"**Detailed Description:**\\n adsfasdf\\n \",\n \"location\": \"adsgasd\",\n \"aifr:harmNarrative\": \"asdgadsf\",\n \"aifr:submitterRelationship\": \"Affected stakeholder\"\n },\n \"aifr:securityAspect\": {\n \"@type\": \"aifr:SecurityIncident\",\n \"aifr:attackerResources\": [\n \"Training data/feedback control \\u2014 An attacker can modify training data/feedback and/or insert a subset of examples.\"\n ],\n \"aifr:attackerObjectives\": [\n \"Availability breakdown \\u2014 An attacker disrupts the ability of users to obtain access to AI systems or their functionality.\"\n ],\n \"aifr:detectionMethods\": [\n \"User observation\"\n ],\n \"aifr:objectiveContext\": \"asdfasdf\"\n },\n \"aifr:disclosure\": {\n \"@type\": \"aifr:DisclosurePlan\",\n \"aifr:intent\": \"No\"\n },\n \"aifr:raw\": {\n \"Submitter Relationship\": \"Affected stakeholder\",\n \"Submitter_Relationship_Other\": \"\",\n \"Incident Location(s)\": \"adsgasd\",\n \"Harm Narrative\": \"asdgadsf\",\n \"Attacker Resources\": [\n \"Training data/feedback control \\u2014 An attacker can modify training data/feedback and/or insert a subset of examples.\"\n ],\n \"Attacker Objectives\": [\n \"Availability breakdown \\u2014 An attacker disrupts the ability of users to obtain access to AI systems or their functionality.\"\n ],\n \"Objective Context\": \"asdfasdf\",\n \"Detection\": [\n \"User observation\"\n ],\n \"Detection_Other\": \"\",\n \"Disclosure Intent\": \"No\",\n \"Disclosure Timeline\": null,\n \"Disclosure Channels\": [],\n \"Disclosure_Channels_Other\": \"\",\n \"Embargo Request\": \"\",\n \"Report Types\": [\n \"Real-World Incidents\",\n \"Malign Actor\",\n \"Security Incident Report\"\n ],\n \"Reporter ID\": \"asdfasdf\",\n \"Session ID\": \"asdfasdf\",\n \"Flaw Timestamp Start\": \"2025-08-19\",\n \"Systems\": [\n \"GPT-3.5-Turbo\"\n ],\n \"Incident Description\": \"**Detailed Description:**\\n adsfasdf\\n \",\n \"Incident Description - Detailed\": \"adsfasdf\",\n \"Potential Policy Violation\": \"asdgasdg\",\n \"Prevalence\": \"Rare\",\n \"Severity\": \"Medium\",\n \"Impacts\": [\n \"Environmental\"\n ],\n \"Impacts_Other\": \"\",\n \"CSAM Related\": null,\n \"Specific Harm Types\": [\n \"Excessive landfill - Excessive disposal of electrical or electronic equipment leading to ecological/biodiversity damage, and disrupting the livelihoods and eroding the rights of local communities\"\n ],\n \"Impacted Stakeholder(s)\": [\n \"Users\"\n ],\n \"Risk Source(s)\": {\n \"Responsible Factors\": [\n \"Tool outputs/external inputs\"\n ],\n \"Responsible Factors Subcategories\": {\n \"Tool outputs/external inputs\": [\n \"Resource overload/accessibility issues\"\n ]\n },\n \"Responsible Factors Context\": \"fasdfasdf\"\n },\n \"Context Info\": \"adfasdf\",\n \"Proof-of-Concept Exploit\": \"asdfasdf\",\n \"Report ID\": \"ef6fa01c-fccd-4430-b81f-d8953d3f70f0\",\n \"Submission Timestamp\": \"2025-08-19T16:16:13.583287\"\n },\n \"aifr:riskSource\": {\n \"@type\": \"aifr:RiskSourceAnalysis\",\n \"aifr:responsibleFactors\": [\n \"Tool outputs/external inputs\"\n ],\n \"aifr:responsibleFactorsSubcategories\": {\n \"Tool outputs/external inputs\": [\n \"Resource overload/accessibility issues\"\n ]\n },\n \"aifr:responsibleFactorsContext\": \"fasdfasdf\"\n },\n \"aifr:contextInfo\": \"adfasdf\",\n \"aifr:submissionTimestamp\": \"2025-08-19T16:16:13.583287\"\n}"
|
54559334-8e48-49f0-9052-ee43ea796b4a
|
Unknown
|
["Real-World Incidents"]
|
asdfasdf
|
2025-08-20T20:46:27.820620
|
{"Submitter Relationship": "Affected stakeholder", "Submitter_Relationship_Other": "", "Incident Location(s)": "asdfasdf", "Harm Narrative": "asdfasdf", "Disclosure Intent": "No", "Disclosure Timeline": null, "Disclosure Channels": [], "Disclosure_Channels_Other": "", "Embargo Request": "", "Report Types": ["Real-World Incidents"], "Reporter ID": "asdfasdf", "Session ID": "asdfasdf", "Flaw Timestamp Start": "2025-08-20", "Systems": ["GPT-4"], "Incident Description": "**Detailed Description:**\n adfasdf\n ", "Incident Description - Detailed": "adfasdf", "Potential Policy Violation": "asdfasdf", "Prevalence": "Rare", "Severity": "Medium", "Impacts": ["Physical"], "Impacts_Other": "", "CSAM Related": null, "Specific Harm Types": ["Bodily injury - Physical pain, injury, illness, or disease suffered by an individual or group due to the malfunction, use or misuse of a technology system"], "Impacted Stakeholder(s)": ["Developers"], "Risk Source(s)": {"Responsible Factors": ["Multi-agent interactions"], "Responsible Factors Subcategories": {"Multi-agent interactions": ["Agent communication poisoning"]}, "Responsible Factors Context": "asdfasdf"}, "Context Info": "asdfasdf", "Proof-of-Concept Exploit": "asdfasd", "Report ID": "54559334-8e48-49f0-9052-ee43ea796b4a", "Submission Timestamp": "2025-08-20T20:46:25.714873"}
|
"{\n \"@context\": [\n \"https://schema.org/\",\n {\n \"aifr\": \"https://aiflawreports.org/schema/\",\n \"aiSystem\": \"aifr:aiSystem\",\n \"severity\": \"aifr:severity\",\n \"prevalence\": \"aifr:prevalence\",\n \"impacts\": \"aifr:impacts\",\n \"reportType\": \"aifr:reportType\",\n \"riskSource\": \"aifr:riskSource\",\n \"contextInfo\": \"aifr:contextInfo\"\n }\n ],\n \"@type\": \"aifr:AIFlawReport\",\n \"@id\": \"https://aiflawreports.org/reports/54559334-8e48-49f0-9052-ee43ea796b4a\",\n \"name\": \"AI Flaw Report: GPT-4\",\n \"description\": \"adfasdf\\n \",\n \"aiSystem\": [\n {\n \"@type\": \"schema:SoftwareApplication\",\n \"@id\": \"https://aiflawreports.org/systems/GPT-4\",\n \"name\": \"GPT-4\",\n \"version\": \"\"\n }\n ],\n \"severity\": \"Medium\",\n \"prevalence\": \"Rare\",\n \"impacts\": [\n \"Physical\"\n ],\n \"reportType\": [\n \"Real-World Incidents\"\n ],\n \"dateCreated\": \"2025-08-20T20:46:25.851022+00:00\",\n \"identifier\": \"54559334-8e48-49f0-9052-ee43ea796b4a\",\n \"aifr:policyViolation\": \"asdfasdf\",\n \"author\": {\n \"@type\": \"schema:Person\",\n \"identifier\": \"asdfasdf\"\n },\n \"aifr:sessionId\": \"asdfasdf\",\n \"aifr:flawTimestamp\": \"2025-08-20\",\n \"aifr:impactedStakeholders\": [\n \"Developers\"\n ],\n \"aifr:specificHarmTypes\": [\n \"Bodily injury - Physical pain, injury, illness, or disease suffered by an individual or group due to the malfunction, use or misuse of a technology system\"\n ],\n \"aifr:incident\": {\n \"@type\": \"aifr:RealWorldIncident\",\n \"description\": \"**Detailed Description:**\\n adfasdf\\n \",\n \"location\": \"asdfasdf\",\n \"aifr:harmNarrative\": \"asdfasdf\",\n \"aifr:submitterRelationship\": \"Affected stakeholder\"\n },\n \"aifr:disclosure\": {\n \"@type\": \"aifr:DisclosurePlan\",\n \"aifr:intent\": \"No\"\n },\n \"aifr:raw\": {\n \"Submitter Relationship\": \"Affected stakeholder\",\n \"Submitter_Relationship_Other\": \"\",\n \"Incident Location(s)\": \"asdfasdf\",\n \"Harm Narrative\": \"asdfasdf\",\n \"Disclosure Intent\": \"No\",\n \"Disclosure Timeline\": null,\n \"Disclosure Channels\": [],\n \"Disclosure_Channels_Other\": \"\",\n \"Embargo Request\": \"\",\n \"Report Types\": [\n \"Real-World Incidents\"\n ],\n \"Reporter ID\": \"asdfasdf\",\n \"Session ID\": \"asdfasdf\",\n \"Flaw Timestamp Start\": \"2025-08-20\",\n \"Systems\": [\n \"GPT-4\"\n ],\n \"Incident Description\": \"**Detailed Description:**\\n adfasdf\\n \",\n \"Incident Description - Detailed\": \"adfasdf\",\n \"Potential Policy Violation\": \"asdfasdf\",\n \"Prevalence\": \"Rare\",\n \"Severity\": \"Medium\",\n \"Impacts\": [\n \"Physical\"\n ],\n \"Impacts_Other\": \"\",\n \"CSAM Related\": null,\n \"Specific Harm Types\": [\n \"Bodily injury - Physical pain, injury, illness, or disease suffered by an individual or group due to the malfunction, use or misuse of a technology system\"\n ],\n \"Impacted Stakeholder(s)\": [\n \"Developers\"\n ],\n \"Risk Source(s)\": {\n \"Responsible Factors\": [\n \"Multi-agent interactions\"\n ],\n \"Responsible Factors Subcategories\": {\n \"Multi-agent interactions\": [\n \"Agent communication poisoning\"\n ]\n },\n \"Responsible Factors Context\": \"asdfasdf\"\n },\n \"Context Info\": \"asdfasdf\",\n \"Proof-of-Concept Exploit\": \"asdfasd\",\n \"Report ID\": \"54559334-8e48-49f0-9052-ee43ea796b4a\",\n \"Submission Timestamp\": \"2025-08-20T20:46:25.714873\"\n },\n \"aifr:riskSource\": {\n \"@type\": \"aifr:RiskSourceAnalysis\",\n \"aifr:responsibleFactors\": [\n \"Multi-agent interactions\"\n ],\n \"aifr:responsibleFactorsSubcategories\": {\n \"Multi-agent interactions\": [\n \"Agent communication poisoning\"\n ]\n },\n \"aifr:responsibleFactorsContext\": \"asdfasdf\"\n },\n \"aifr:contextInfo\": \"asdfasdf\",\n \"aifr:submissionTimestamp\": \"2025-08-20T20:46:25.714873\"\n}"
|
626a9a3f-8550-4059-99ab-ada6393b7c9f
|
Unknown
|
["Real-World Incidents", "Malign Actor", "Security Incident Report"]
|
Basic name
|
2025-08-22T13:07:52.881338
|
{"Submitter Relationship": "Independent observer", "Submitter_Relationship_Other": "", "Incident Location(s)": null, "Harm Narrative": null, "Attacker Resources": [], "Attacker Objectives": [], "Objective Context": "", "Detection": [], "Detection_Other": "", "Disclosure Intent": "Yes", "Disclosure Timeline": "Short-term (1-30 days)", "Disclosure Channels": ["Academic paper", "Media outlet"], "Disclosure_Channels_Other": "", "Embargo Request": "Embargo details", "Report Types": ["Real-World Incidents", "Malign Actor", "Security Incident Report"], "Reporter ID": "Basic name", "Session ID": "session ID", "Flaw Timestamp Start": "2025-08-29", "Systems": ["GPT-4.5-Preview", "GPT-3.5-Turbo", "Other"], "Incident Description": "**Detailed Description:**\n Incident description\n ", "Incident Description - Detailed": "Incident description", "Potential Policy Violation": "Policy violations!", "Prevalence": "Rare", "Severity": "Significant", "Impacts": ["Environmental", "Fundamental rights"], "Impacts_Other": "", "CSAM Related": null, "Specific Harm Types": ["Excessive energy consumption - Excessive energy use, leading to energy bottlenecks and shortages for communities, organisations, and businesses", "Excessive water consumption - Excessive use of water to cool data centres and for other purposes, leading to water restrictions or shortages for local communities or businesses"], "Impacted Stakeholder(s)": ["General Public", "Administrators"], "Risk Source(s)": {"Responsible Factors": ["Tool outputs/external inputs", "System prompt"], "Responsible Factors Subcategories": {"Tool outputs/external inputs": ["Reliability issues (e.g., false information)"]}, "Responsible Factors Context": null}, "Context Info": null, "Proof-of-Concept Exploit": null, "Report ID": "626a9a3f-8550-4059-99ab-ada6393b7c9f", "Submission Timestamp": "2025-08-22T13:07:51.007306"}
|
"{\n \"@context\": [\n \"https://schema.org/\",\n {\n \"aifr\": \"https://aiflawreports.org/schema/\",\n \"aiSystem\": \"aifr:aiSystem\",\n \"severity\": \"aifr:severity\",\n \"prevalence\": \"aifr:prevalence\",\n \"impacts\": \"aifr:impacts\",\n \"reportType\": \"aifr:reportType\",\n \"riskSource\": \"aifr:riskSource\",\n \"contextInfo\": \"aifr:contextInfo\"\n }\n ],\n \"@type\": \"aifr:AIFlawReport\",\n \"@id\": \"https://aiflawreports.org/reports/626a9a3f-8550-4059-99ab-ada6393b7c9f\",\n \"name\": \"AI Flaw Report: GPT-4.5-Preview, GPT-3.5-Turbo, Other\",\n \"description\": \"Incident description\\n \",\n \"aiSystem\": [\n {\n \"@type\": \"schema:SoftwareApplication\",\n \"@id\": \"https://aiflawreports.org/systems/GPT-4.5-Preview\",\n \"name\": \"GPT-4.5-Preview\",\n \"version\": \"\"\n },\n {\n \"@type\": \"schema:SoftwareApplication\",\n \"@id\": \"https://aiflawreports.org/systems/GPT-3.5-Turbo\",\n \"name\": \"GPT-3.5-Turbo\",\n \"version\": \"\"\n },\n {\n \"@type\": \"schema:SoftwareApplication\",\n \"@id\": \"https://aiflawreports.org/systems/Other\",\n \"name\": \"Other\",\n \"version\": \"\"\n }\n ],\n \"severity\": \"Significant\",\n \"prevalence\": \"Rare\",\n \"impacts\": [\n \"Environmental\",\n \"Fundamental rights\"\n ],\n \"reportType\": [\n \"Real-World Incidents\",\n \"Malign Actor\",\n \"Security Incident Report\"\n ],\n \"dateCreated\": \"2025-08-22T13:07:51.142471+00:00\",\n \"identifier\": \"626a9a3f-8550-4059-99ab-ada6393b7c9f\",\n \"aifr:policyViolation\": \"Policy violations!\",\n \"author\": {\n \"@type\": \"schema:Person\",\n \"identifier\": \"Basic name\"\n },\n \"aifr:sessionId\": \"session ID\",\n \"aifr:flawTimestamp\": \"2025-08-29\",\n \"aifr:impactedStakeholders\": [\n \"General Public\",\n \"Administrators\"\n ],\n \"aifr:specificHarmTypes\": [\n \"Excessive energy consumption - Excessive energy use, leading to energy bottlenecks and shortages for communities, organisations, and businesses\",\n \"Excessive water consumption - Excessive use of water to cool data centres and for other purposes, leading to water restrictions or shortages for local communities or businesses\"\n ],\n \"aifr:incident\": {\n \"@type\": \"aifr:RealWorldIncident\",\n \"description\": \"**Detailed Description:**\\n Incident description\\n \",\n \"location\": null,\n \"aifr:harmNarrative\": null,\n \"aifr:submitterRelationship\": \"Independent observer\"\n },\n \"aifr:securityAspect\": {\n \"@type\": \"aifr:SecurityIncident\",\n \"aifr:attackerResources\": [],\n \"aifr:attackerObjectives\": [],\n \"aifr:detectionMethods\": []\n },\n \"aifr:disclosure\": {\n \"@type\": \"aifr:DisclosurePlan\",\n \"aifr:intent\": \"Yes\",\n \"aifr:timeline\": \"Short-term (1-30 days)\",\n \"aifr:channels\": [\n \"Academic paper\",\n \"Media outlet\"\n ],\n \"aifr:embargoRequest\": \"Embargo details\"\n },\n \"aifr:raw\": {\n \"Submitter Relationship\": \"Independent observer\",\n \"Submitter_Relationship_Other\": \"\",\n \"Incident Location(s)\": null,\n \"Harm Narrative\": null,\n \"Attacker Resources\": [],\n \"Attacker Objectives\": [],\n \"Objective Context\": \"\",\n \"Detection\": [],\n \"Detection_Other\": \"\",\n \"Disclosure Intent\": \"Yes\",\n \"Disclosure Timeline\": \"Short-term (1-30 days)\",\n \"Disclosure Channels\": [\n \"Academic paper\",\n \"Media outlet\"\n ],\n \"Disclosure_Channels_Other\": \"\",\n \"Embargo Request\": \"Embargo details\",\n \"Report Types\": [\n \"Real-World Incidents\",\n \"Malign Actor\",\n \"Security Incident Report\"\n ],\n \"Reporter ID\": \"Basic name\",\n \"Session ID\": \"session ID\",\n \"Flaw Timestamp Start\": \"2025-08-29\",\n \"Systems\": [\n \"GPT-4.5-Preview\",\n \"GPT-3.5-Turbo\",\n \"Other\"\n ],\n \"Incident Description\": \"**Detailed Description:**\\n Incident description\\n \",\n \"Incident Description - Detailed\": \"Incident description\",\n \"Potential Policy Violation\": \"Policy violations!\",\n \"Prevalence\": \"Rare\",\n \"Severity\": \"Significant\",\n \"Impacts\": [\n \"Environmental\",\n \"Fundamental rights\"\n ],\n \"Impacts_Other\": \"\",\n \"CSAM Related\": null,\n \"Specific Harm Types\": [\n \"Excessive energy consumption - Excessive energy use, leading to energy bottlenecks and shortages for communities, organisations, and businesses\",\n \"Excessive water consumption - Excessive use of water to cool data centres and for other purposes, leading to water restrictions or shortages for local communities or businesses\"\n ],\n \"Impacted Stakeholder(s)\": [\n \"General Public\",\n \"Administrators\"\n ],\n \"Risk Source(s)\": {\n \"Responsible Factors\": [\n \"Tool outputs/external inputs\",\n \"System prompt\"\n ],\n \"Responsible Factors Subcategories\": {\n \"Tool outputs/external inputs\": [\n \"Reliability issues (e.g., false information)\"\n ]\n },\n \"Responsible Factors Context\": null\n },\n \"Context Info\": null,\n \"Proof-of-Concept Exploit\": null,\n \"Report ID\": \"626a9a3f-8550-4059-99ab-ada6393b7c9f\",\n \"Submission Timestamp\": \"2025-08-22T13:07:51.007306\"\n },\n \"aifr:riskSource\": {\n \"@type\": \"aifr:RiskSourceAnalysis\",\n \"aifr:responsibleFactors\": [\n \"Tool outputs/external inputs\",\n \"System prompt\"\n ],\n \"aifr:responsibleFactorsSubcategories\": {\n \"Tool outputs/external inputs\": [\n \"Reliability issues (e.g., false information)\"\n ]\n },\n \"aifr:responsibleFactorsContext\": null\n },\n \"aifr:submissionTimestamp\": \"2025-08-22T13:07:51.007306\"\n}"
|
49ea94e7-27e6-4716-98a0-208b5cb78118
|
Unknown
|
["Real-World Incidents", "Malign Actor", "Security Incident Report"]
|
zhu.lae@northeastern.edu
|
2025-09-18T21:14:32.391638
|
{"Submitter Relationship": "Independent observer", "Submitter_Relationship_Other": "", "Incident Location(s)": "Global", "Harm Narrative": "The incident demonstrates that seemingly innocuous API parameters can be combined to extract proprietary model information, undermining the security model of \"black box\" API access. This creates precedent for model stealing attacks on production systems.", "Attacker Resources": ["Direct query access \u2014 black-box \u2014 An attacker can query the system\u2014the degree of control can vary substantially (e.g., ability to control temperature, view logits, etc.).", "Application/plugin supply chain control \u2014 An attacker can modify the agent framework, tools and/or services with which a model interacts, such as introducing vulnerabilities into application software or appending malicious text to plugin instructions."], "Attacker Objectives": ["Privacy compromise \u2014 An attacker gains access to sensitive and confidential information, including information about the AI system (e.g., architecture or weights) or sensitive information that the model accesses (e.g., training data, external knowledge databases.", "Integrity violation \u2014 An attacker causes AI systems to perform tasks inadequately or behave undesirably."], "Objective Context": null, "Detection": ["Testing", "External report"], "Detection_Other": "", "Disclosure Intent": "Already Public Knowledge", "Disclosure Timeline": null, "Disclosure Channels": [], "Disclosure_Channels_Other": "", "Embargo Request": "", "Report Types": ["Real-World Incidents", "Malign Actor", "Security Incident Report"], "Uploaded Files": ["2403.06634v2.pdf"], "Uploaded File Paths": ["uploads/2403.06634v2.pdf"], "Reporter ID": "zhu.lae@northeastern.edu", "Session ID": null, "Flaw Timestamp Start": "2024-01-01", "Systems": ["GPT-3.5-Turbo"], "Incident Description": "**Detailed Description:**\n Researchers successfully extracted the complete embedding projection layer (final layer weights) from production language models by exploiting API features that provided logit probabilities and logit bias parameters. The attack recovered precise model architecture details (hidden dimensions) and weight matrices with mean squared error of 10^-4. For under $20 USD, the attack extracted entire projection matrices from OpenAI's ada and babbage models, confirming hidden dimensions of 1024 and 2048 respectively.\n ", "Incident Description - Detailed": "Researchers successfully extracted the complete embedding projection layer (final layer weights) from production language models by exploiting API features that provided logit probabilities and logit bias parameters. The attack recovered precise model architecture details (hidden dimensions) and weight matrices with mean squared error of 10^-4. For under $20 USD, the attack extracted entire projection matrices from OpenAI's ada and babbage models, confirming hidden dimensions of 1024 and 2048 respectively.", "Potential Policy Violation": "This attack violates the implicit expectation that model weights and architecture details remain proprietary. It potentially violates terms of service regarding reverse engineering and extracting confidential information about model internals. The attack undermines the \"black box\" nature that API providers expect to maintain.", "Prevalence": "Occasional", "Severity": "Significant", "Impacts": ["Economic/property"], "Impacts_Other": "", "CSAM Related": null, "Specific Harm Types": [], "Impacted Stakeholder(s)": ["Developers", "Organizations"], "Risk Source(s)": {"Responsible Factors": ["System prompt", "Supply chain weaknesses (e.g., software libraries and hardware)"], "Responsible Factors Subcategories": {}, "Responsible Factors Context": "System prompt: API design exposing logit bias and probabilities\nSupply chain weakness: API parameter combinations"}, "Context Info": "Attack requires API access with logit bias and logprobs parameters. Uses mathematical techniques including SVD to extract hidden dimensions from logit vectors across multiple queries with varying bias parameters.", "Proof-of-Concept Exploit": "The paper provides complete mathematical methodology and algorithms for dimension extraction and weight matrix recovery. Code available in supplementary materials.", "Report ID": "49ea94e7-27e6-4716-98a0-208b5cb78118", "Submission Timestamp": "2025-09-18T21:10:01.829968"}
|
"{\n \"@context\": [\n \"https://schema.org/\",\n {\n \"aifr\": \"https://aiflawreports.org/schema/\",\n \"aiSystem\": \"aifr:aiSystem\",\n \"severity\": \"aifr:severity\",\n \"prevalence\": \"aifr:prevalence\",\n \"impacts\": \"aifr:impacts\",\n \"reportType\": \"aifr:reportType\",\n \"riskSource\": \"aifr:riskSource\",\n \"contextInfo\": \"aifr:contextInfo\"\n }\n ],\n \"@type\": \"aifr:AIFlawReport\",\n \"@id\": \"https://aiflawreports.org/reports/49ea94e7-27e6-4716-98a0-208b5cb78118\",\n \"name\": \"AI Flaw Report: GPT-3.5-Turbo\",\n \"description\": \"Researchers successfully extracted the complete embedding projection layer (final layer weights) from production language models by exploiting API features that provided logit probabilities and logit bias parameters. The attack recovered precise model architecture details (hidden dimensions) and weight matrices with mean squared error of 10^-4. For under $20 USD, the attack extracted entire projection matrices from OpenAI's ada and babbage models, confirming hidden dimensions of 1024 and 2048 respectively.\\n \",\n \"aiSystem\": [\n {\n \"@type\": \"schema:SoftwareApplication\",\n \"@id\": \"https://aiflawreports.org/systems/GPT-3.5-Turbo\",\n \"name\": \"GPT-3.5-Turbo\",\n \"version\": \"\"\n }\n ],\n \"severity\": \"Significant\",\n \"prevalence\": \"Occasional\",\n \"impacts\": [\n \"Economic/property\"\n ],\n \"reportType\": [\n \"Real-World Incidents\",\n \"Malign Actor\",\n \"Security Incident Report\"\n ],\n \"dateCreated\": \"2025-09-18T21:14:30.617113+00:00\",\n \"identifier\": \"49ea94e7-27e6-4716-98a0-208b5cb78118\",\n \"aifr:policyViolation\": \"This attack violates the implicit expectation that model weights and architecture details remain proprietary. It potentially violates terms of service regarding reverse engineering and extracting confidential information about model internals. The attack undermines the \\\"black box\\\" nature that API providers expect to maintain.\",\n \"author\": {\n \"@type\": \"schema:Person\",\n \"identifier\": \"zhu.lae@northeastern.edu\"\n },\n \"aifr:flawTimestamp\": \"2024-01-01\",\n \"aifr:impactedStakeholders\": [\n \"Developers\",\n \"Organizations\"\n ],\n \"aifr:incident\": {\n \"@type\": \"aifr:RealWorldIncident\",\n \"description\": \"**Detailed Description:**\\n Researchers successfully extracted the complete embedding projection layer (final layer weights) from production language models by exploiting API features that provided logit probabilities and logit bias parameters. The attack recovered precise model architecture details (hidden dimensions) and weight matrices with mean squared error of 10^-4. For under $20 USD, the attack extracted entire projection matrices from OpenAI's ada and babbage models, confirming hidden dimensions of 1024 and 2048 respectively.\\n \",\n \"location\": \"Global\",\n \"aifr:harmNarrative\": \"The incident demonstrates that seemingly innocuous API parameters can be combined to extract proprietary model information, undermining the security model of \\\"black box\\\" API access. This creates precedent for model stealing attacks on production systems.\",\n \"aifr:submitterRelationship\": \"Independent observer\"\n },\n \"aifr:securityAspect\": {\n \"@type\": \"aifr:SecurityIncident\",\n \"aifr:attackerResources\": [\n \"Direct query access \\u2014 black-box \\u2014 An attacker can query the system\\u2014the degree of control can vary substantially (e.g., ability to control temperature, view logits, etc.).\",\n \"Application/plugin supply chain control \\u2014 An attacker can modify the agent framework, tools and/or services with which a model interacts, such as introducing vulnerabilities into application software or appending malicious text to plugin instructions.\"\n ],\n \"aifr:attackerObjectives\": [\n \"Privacy compromise \\u2014 An attacker gains access to sensitive and confidential information, including information about the AI system (e.g., architecture or weights) or sensitive information that the model accesses (e.g., training data, external knowledge databases.\",\n \"Integrity violation \\u2014 An attacker causes AI systems to perform tasks inadequately or behave undesirably.\"\n ],\n \"aifr:detectionMethods\": [\n \"Testing\",\n \"External report\"\n ]\n },\n \"aifr:disclosure\": {\n \"@type\": \"aifr:DisclosurePlan\",\n \"aifr:intent\": \"Already Public Knowledge\"\n },\n \"aifr:raw\": {\n \"Submitter Relationship\": \"Independent observer\",\n \"Submitter_Relationship_Other\": \"\",\n \"Incident Location(s)\": \"Global\",\n \"Harm Narrative\": \"The incident demonstrates that seemingly innocuous API parameters can be combined to extract proprietary model information, undermining the security model of \\\"black box\\\" API access. This creates precedent for model stealing attacks on production systems.\",\n \"Attacker Resources\": [\n \"Direct query access \\u2014 black-box \\u2014 An attacker can query the system\\u2014the degree of control can vary substantially (e.g., ability to control temperature, view logits, etc.).\",\n \"Application/plugin supply chain control \\u2014 An attacker can modify the agent framework, tools and/or services with which a model interacts, such as introducing vulnerabilities into application software or appending malicious text to plugin instructions.\"\n ],\n \"Attacker Objectives\": [\n \"Privacy compromise \\u2014 An attacker gains access to sensitive and confidential information, including information about the AI system (e.g., architecture or weights) or sensitive information that the model accesses (e.g., training data, external knowledge databases.\",\n \"Integrity violation \\u2014 An attacker causes AI systems to perform tasks inadequately or behave undesirably.\"\n ],\n \"Objective Context\": null,\n \"Detection\": [\n \"Testing\",\n \"External report\"\n ],\n \"Detection_Other\": \"\",\n \"Disclosure Intent\": \"Already Public Knowledge\",\n \"Disclosure Timeline\": null,\n \"Disclosure Channels\": [],\n \"Disclosure_Channels_Other\": \"\",\n \"Embargo Request\": \"\",\n \"Report Types\": [\n \"Real-World Incidents\",\n \"Malign Actor\",\n \"Security Incident Report\"\n ],\n \"Uploaded Files\": [\n \"2403.06634v2.pdf\"\n ],\n \"Uploaded File Paths\": [\n \"uploads/2403.06634v2.pdf\"\n ],\n \"Reporter ID\": \"zhu.lae@northeastern.edu\",\n \"Session ID\": null,\n \"Flaw Timestamp Start\": \"2024-01-01\",\n \"Systems\": [\n \"GPT-3.5-Turbo\"\n ],\n \"Incident Description\": \"**Detailed Description:**\\n Researchers successfully extracted the complete embedding projection layer (final layer weights) from production language models by exploiting API features that provided logit probabilities and logit bias parameters. The attack recovered precise model architecture details (hidden dimensions) and weight matrices with mean squared error of 10^-4. For under $20 USD, the attack extracted entire projection matrices from OpenAI's ada and babbage models, confirming hidden dimensions of 1024 and 2048 respectively.\\n \",\n \"Incident Description - Detailed\": \"Researchers successfully extracted the complete embedding projection layer (final layer weights) from production language models by exploiting API features that provided logit probabilities and logit bias parameters. The attack recovered precise model architecture details (hidden dimensions) and weight matrices with mean squared error of 10^-4. For under $20 USD, the attack extracted entire projection matrices from OpenAI's ada and babbage models, confirming hidden dimensions of 1024 and 2048 respectively.\",\n \"Potential Policy Violation\": \"This attack violates the implicit expectation that model weights and architecture details remain proprietary. It potentially violates terms of service regarding reverse engineering and extracting confidential information about model internals. The attack undermines the \\\"black box\\\" nature that API providers expect to maintain.\",\n \"Prevalence\": \"Occasional\",\n \"Severity\": \"Significant\",\n \"Impacts\": [\n \"Economic/property\"\n ],\n \"Impacts_Other\": \"\",\n \"CSAM Related\": null,\n \"Specific Harm Types\": [],\n \"Impacted Stakeholder(s)\": [\n \"Developers\",\n \"Organizations\"\n ],\n \"Risk Source(s)\": {\n \"Responsible Factors\": [\n \"System prompt\",\n \"Supply chain weaknesses (e.g., software libraries and hardware)\"\n ],\n \"Responsible Factors Subcategories\": {},\n \"Responsible Factors Context\": \"System prompt: API design exposing logit bias and probabilities\\nSupply chain weakness: API parameter combinations\"\n },\n \"Context Info\": \"Attack requires API access with logit bias and logprobs parameters. Uses mathematical techniques including SVD to extract hidden dimensions from logit vectors across multiple queries with varying bias parameters.\",\n \"Proof-of-Concept Exploit\": \"The paper provides complete mathematical methodology and algorithms for dimension extraction and weight matrix recovery. Code available in supplementary materials.\",\n \"Report ID\": \"49ea94e7-27e6-4716-98a0-208b5cb78118\",\n \"Submission Timestamp\": \"2025-09-18T21:10:01.829968\"\n },\n \"aifr:riskSource\": {\n \"@type\": \"aifr:RiskSourceAnalysis\",\n \"aifr:responsibleFactors\": [\n \"System prompt\",\n \"Supply chain weaknesses (e.g., software libraries and hardware)\"\n ],\n \"aifr:responsibleFactorsSubcategories\": {},\n \"aifr:responsibleFactorsContext\": \"System prompt: API design exposing logit bias and probabilities\\nSupply chain weakness: API parameter combinations\"\n },\n \"aifr:contextInfo\": \"Attack requires API access with logit bias and logprobs parameters. Uses mathematical techniques including SVD to extract hidden dimensions from logit vectors across multiple queries with varying bias parameters.\",\n \"aifr:submissionTimestamp\": \"2025-09-18T21:10:01.829968\"\n}"
|
2a8ce1b3-c6e3-48d3-9ab7-993448631801
|
Unknown
|
["Hazard Report"]
|
asdfasdf
|
2025-09-18T17:52:08.235137
|
{"Statistical Argument with Examples": "asdfasdf", "Disclosure Intent": "No", "Disclosure Timeline": null, "Disclosure Channels": [], "Disclosure_Channels_Other": "", "Embargo Request": "", "Report Types": ["Hazard Report"], "Reporter ID": "asdfasdf", "Session ID": null, "Flaw Timestamp Start": "2025-09-17", "Systems": ["o3-mini"], "Flaw Description": "**Detailed Description:**\n adfasdf\n ", "Flaw Description - Detailed": "adfasdf", "Policy Violation": "asdfasdf", "Prevalence": "Rare", "Severity": "Low", "Impacts": ["Societal and Cultural"], "Impacts_Other": "", "CSAM Related": null, "Specific Harm Types": ["Cheating/plagiarism - Use of another person's or group's words or ideas without consent and/or acknowledgement"], "Impacted Stakeholder(s)": ["Organizations"], "Risk Source(s)": {"Responsible Factors": [], "Responsible Factors Subcategories": {}, "Responsible Factors Context": ""}, "Context Info": "asdfasdf", "Proof-of-Concept Exploit": "asdfadsf", "Report ID": "2a8ce1b3-c6e3-48d3-9ab7-993448631801", "Submission Timestamp": "2025-09-18T17:51:59.671047"}
|
"{\n \"@context\": [\n \"https://schema.org/\",\n {\n \"aifr\": \"https://aiflawreports.org/schema/\",\n \"aiSystem\": \"aifr:aiSystem\",\n \"severity\": \"aifr:severity\",\n \"prevalence\": \"aifr:prevalence\",\n \"impacts\": \"aifr:impacts\",\n \"reportType\": \"aifr:reportType\",\n \"riskSource\": \"aifr:riskSource\",\n \"contextInfo\": \"aifr:contextInfo\"\n }\n ],\n \"@type\": \"aifr:AIFlawReport\",\n \"@id\": \"https://aiflawreports.org/reports/2a8ce1b3-c6e3-48d3-9ab7-993448631801\",\n \"name\": \"AI Flaw Report: o3-mini\",\n \"description\": \"adfasdf\\n \",\n \"aiSystem\": [\n {\n \"@type\": \"schema:SoftwareApplication\",\n \"@id\": \"https://aiflawreports.org/systems/o3-mini\",\n \"name\": \"o3-mini\",\n \"version\": \"\"\n }\n ],\n \"severity\": \"Low\",\n \"prevalence\": \"Rare\",\n \"impacts\": [\n \"Societal and Cultural\"\n ],\n \"reportType\": [\n \"Hazard Report\"\n ],\n \"dateCreated\": \"2025-09-18T21:52:06.581943+00:00\",\n \"identifier\": \"2a8ce1b3-c6e3-48d3-9ab7-993448631801\",\n \"aifr:policyViolation\": \"asdfasdf\",\n \"author\": {\n \"@type\": \"schema:Person\",\n \"identifier\": \"asdfasdf\"\n },\n \"aifr:flawTimestamp\": \"2025-09-17\",\n \"aifr:impactedStakeholders\": [\n \"Organizations\"\n ],\n \"aifr:specificHarmTypes\": [\n \"Cheating/plagiarism - Use of another person's or group's words or ideas without consent and/or acknowledgement\"\n ],\n \"aifr:hazard\": {\n \"@type\": \"aifr:Hazard\",\n \"aifr:statisticalArgument\": \"asdfasdf\"\n },\n \"aifr:disclosure\": {\n \"@type\": \"aifr:DisclosurePlan\",\n \"aifr:intent\": \"No\"\n },\n \"aifr:raw\": {\n \"Statistical Argument with Examples\": \"asdfasdf\",\n \"Disclosure Intent\": \"No\",\n \"Disclosure Timeline\": null,\n \"Disclosure Channels\": [],\n \"Disclosure_Channels_Other\": \"\",\n \"Embargo Request\": \"\",\n \"Report Types\": [\n \"Hazard Report\"\n ],\n \"Reporter ID\": \"asdfasdf\",\n \"Session ID\": null,\n \"Flaw Timestamp Start\": \"2025-09-17\",\n \"Systems\": [\n \"o3-mini\"\n ],\n \"Flaw Description\": \"**Detailed Description:**\\n adfasdf\\n \",\n \"Flaw Description - Detailed\": \"adfasdf\",\n \"Policy Violation\": \"asdfasdf\",\n \"Prevalence\": \"Rare\",\n \"Severity\": \"Low\",\n \"Impacts\": [\n \"Societal and Cultural\"\n ],\n \"Impacts_Other\": \"\",\n \"CSAM Related\": null,\n \"Specific Harm Types\": [\n \"Cheating/plagiarism - Use of another person's or group's words or ideas without consent and/or acknowledgement\"\n ],\n \"Impacted Stakeholder(s)\": [\n \"Organizations\"\n ],\n \"Risk Source(s)\": {\n \"Responsible Factors\": [],\n \"Responsible Factors Subcategories\": {},\n \"Responsible Factors Context\": \"\"\n },\n \"Context Info\": \"asdfasdf\",\n \"Proof-of-Concept Exploit\": \"asdfadsf\",\n \"Report ID\": \"2a8ce1b3-c6e3-48d3-9ab7-993448631801\",\n \"Submission Timestamp\": \"2025-09-18T17:51:59.671047\"\n },\n \"aifr:riskSource\": {\n \"@type\": \"aifr:RiskSourceAnalysis\",\n \"aifr:responsibleFactors\": [],\n \"aifr:responsibleFactorsSubcategories\": {},\n \"aifr:responsibleFactorsContext\": \"\"\n },\n \"aifr:contextInfo\": \"asdfasdf\",\n \"aifr:submissionTimestamp\": \"2025-09-18T17:51:59.671047\"\n}"
|
d71f5b3f-63f7-4ec1-bc51-015e80f8ef24
|
Unknown
|
["Hazard Report"]
|
adfasdf
|
2025-09-18T17:58:44.545691
|
{"Reporter ID": "adfasdf", "Session ID": "asdfasd", "Flaw Timestamp Start": "2025-09-18", "Systems": ["GPT-3.5-Turbo"], "Flaw Description": "**Detailed Description:**\n fasdfa\n ", "Flaw Description - Detailed": "fasdfa", "Policy Violation": "asdfadsf", "Prevalence": "Rare", "Severity": "Low", "Impacts": ["Financial and Business"], "Impacts_Other": "", "CSAM Related": null, "Specific Harm Types": ["Increased competition - The inappropriate or unethical use of technology to gain market share"], "Impacted Stakeholder(s)": ["Organizations"], "Risk Source(s)": {"Responsible Factors": ["Feedback"], "Responsible Factors Subcategories": {"Feedback": ["Poisoning"]}, "Responsible Factors Context": "adfasdf"}, "Context Info": "asdfasdf", "Proof-of-Concept Exploit": "asdfasdf", "Disclosure Intent": "No", "Disclosure Timeline": null, "Disclosure Channels": [], "Disclosure_Channels_Other": "", "Embargo Request": "", "Statistical Argument with Examples": "afsdfasf", "Report Types": ["Hazard Report"], "Report ID": "d71f5b3f-63f7-4ec1-bc51-015e80f8ef24", "Submission Timestamp": "2025-09-18T17:58:26.594848"}
|
"{\n \"@context\": [\n \"https://schema.org/\",\n {\n \"aifr\": \"https://aiflawreports.org/schema/\",\n \"aiSystem\": \"aifr:aiSystem\",\n \"severity\": \"aifr:severity\",\n \"prevalence\": \"aifr:prevalence\",\n \"impacts\": \"aifr:impacts\",\n \"reportType\": \"aifr:reportType\",\n \"riskSource\": \"aifr:riskSource\",\n \"contextInfo\": \"aifr:contextInfo\"\n }\n ],\n \"@type\": \"aifr:AIFlawReport\",\n \"@id\": \"https://aiflawreports.org/reports/d71f5b3f-63f7-4ec1-bc51-015e80f8ef24\",\n \"name\": \"AI Flaw Report: GPT-3.5-Turbo\",\n \"description\": \"fasdfa\\n \",\n \"aiSystem\": [\n {\n \"@type\": \"schema:SoftwareApplication\",\n \"@id\": \"https://aiflawreports.org/systems/GPT-3.5-Turbo\",\n \"name\": \"GPT-3.5-Turbo\",\n \"version\": \"\"\n }\n ],\n \"severity\": \"Low\",\n \"prevalence\": \"Rare\",\n \"impacts\": [\n \"Financial and Business\"\n ],\n \"reportType\": [\n \"Hazard Report\"\n ],\n \"dateCreated\": \"2025-09-18T21:58:43.113800+00:00\",\n \"identifier\": \"d71f5b3f-63f7-4ec1-bc51-015e80f8ef24\",\n \"aifr:policyViolation\": \"asdfadsf\",\n \"author\": {\n \"@type\": \"schema:Person\",\n \"identifier\": \"adfasdf\"\n },\n \"aifr:sessionId\": \"asdfasd\",\n \"aifr:flawTimestamp\": \"2025-09-18\",\n \"aifr:impactedStakeholders\": [\n \"Organizations\"\n ],\n \"aifr:specificHarmTypes\": [\n \"Increased competition - The inappropriate or unethical use of technology to gain market share\"\n ],\n \"aifr:hazard\": {\n \"@type\": \"aifr:Hazard\",\n \"aifr:statisticalArgument\": \"afsdfasf\"\n },\n \"aifr:disclosure\": {\n \"@type\": \"aifr:DisclosurePlan\",\n \"aifr:intent\": \"No\"\n },\n \"aifr:raw\": {\n \"Reporter ID\": \"adfasdf\",\n \"Session ID\": \"asdfasd\",\n \"Flaw Timestamp Start\": \"2025-09-18\",\n \"Systems\": [\n \"GPT-3.5-Turbo\"\n ],\n \"Flaw Description\": \"**Detailed Description:**\\n fasdfa\\n \",\n \"Flaw Description - Detailed\": \"fasdfa\",\n \"Policy Violation\": \"asdfadsf\",\n \"Prevalence\": \"Rare\",\n \"Severity\": \"Low\",\n \"Impacts\": [\n \"Financial and Business\"\n ],\n \"Impacts_Other\": \"\",\n \"CSAM Related\": null,\n \"Specific Harm Types\": [\n \"Increased competition - The inappropriate or unethical use of technology to gain market share\"\n ],\n \"Impacted Stakeholder(s)\": [\n \"Organizations\"\n ],\n \"Risk Source(s)\": {\n \"Responsible Factors\": [\n \"Feedback\"\n ],\n \"Responsible Factors Subcategories\": {\n \"Feedback\": [\n \"Poisoning\"\n ]\n },\n \"Responsible Factors Context\": \"adfasdf\"\n },\n \"Context Info\": \"asdfasdf\",\n \"Proof-of-Concept Exploit\": \"asdfasdf\",\n \"Disclosure Intent\": \"No\",\n \"Disclosure Timeline\": null,\n \"Disclosure Channels\": [],\n \"Disclosure_Channels_Other\": \"\",\n \"Embargo Request\": \"\",\n \"Statistical Argument with Examples\": \"afsdfasf\",\n \"Report Types\": [\n \"Hazard Report\"\n ],\n \"Report ID\": \"d71f5b3f-63f7-4ec1-bc51-015e80f8ef24\",\n \"Submission Timestamp\": \"2025-09-18T17:58:26.594848\"\n },\n \"aifr:riskSource\": {\n \"@type\": \"aifr:RiskSourceAnalysis\",\n \"aifr:responsibleFactors\": [\n \"Feedback\"\n ],\n \"aifr:responsibleFactorsSubcategories\": {\n \"Feedback\": [\n \"Poisoning\"\n ]\n },\n \"aifr:responsibleFactorsContext\": \"adfasdf\"\n },\n \"aifr:contextInfo\": \"asdfasdf\",\n \"aifr:submissionTimestamp\": \"2025-09-18T17:58:26.594848\"\n}"
|
67120781-6f5a-4bdf-973b-e2355161a7bd
|
Unknown
|
["Real-World Incidents", "Malign Actor", "Security Incident Report"]
|
zhu.lae@northeastern.edu
|
2025-09-18T18:15:11.234541
|
{"Reporter ID": "zhu.lae@northeastern.edu", "Session ID": null, "Flaw Timestamp Start": "2024-01-01", "Systems": ["GPT-3.5-Turbo"], "Incident Description": "**Detailed Description:**\n I extracted the complete embedding projection layer (final layer weights) from production language models by exploiting API features that provided logit probabilities and logit bias parameters. My attack recovered precise model architecture details (hidden dimensions) and weight matrices with mean squared error of 10^-4. I extracted entire projection matrices from OpenAI's ada and babbage models, confirming hidden dimensions of 1024 and 2048 respectively.\n ", "Incident Description - Detailed": "I extracted the complete embedding projection layer (final layer weights) from production language models by exploiting API features that provided logit probabilities and logit bias parameters. My attack recovered precise model architecture details (hidden dimensions) and weight matrices with mean squared error of 10^-4. I extracted entire projection matrices from OpenAI's ada and babbage models, confirming hidden dimensions of 1024 and 2048 respectively.", "Potential Policy Violation": "This attack violates the implicit expectation that model weights and architecture details remain proprietary. It potentially violates terms of service regarding reverse engineering and extracting confidential information about model internals. The attack undermines the \"black box\" nature that API providers expect to maintain.", "Prevalence": "Occasional", "Severity": "Significant", "Impacts": ["Economic/property", "Other"], "Impacts_Other": "Privacy", "CSAM Related": null, "Specific Harm Types": [], "Impacted Stakeholder(s)": ["Developers", "Organizations"], "Risk Source(s)": {"Responsible Factors": ["System prompt", "Supply chain weaknesses (e.g., software libraries and hardware)"], "Responsible Factors Subcategories": {}, "Responsible Factors Context": "System prompt: API design exposing logit bias and probabilities\nSupply chain weakness: API parameter combinations"}, "Context Info": "Attack requires API access with logit bias and logprobs parameters. Uses mathematical techniques including SVD to extract hidden dimensions from logit vectors across multiple queries with varying bias parameters.", "Proof-of-Concept Exploit": "I provide complete mathematical methodology and algorithms for dimension extraction and weight matrix recovery. Code available in my supplementary materials.", "Disclosure Intent": "Yes", "Disclosure Timeline": "Short-term (1-30 days)", "Disclosure Channels": ["Academic paper"], "Disclosure_Channels_Other": "", "Embargo Request": null, "Submitter Relationship": "Independent observer", "Submitter_Relationship_Other": "", "Incident Location(s)": "Global", "Harm Narrative": "The incident demonstrates that seemingly innocuous API parameters can be combined to extract proprietary model information, undermining the security model of \"black box\" API access. This creates precedent for model stealing attacks on production systems.", "Attacker Resources": ["Direct query access \u2014 black-box \u2014 An attacker can query the system\u2014the degree of control can vary substantially (e.g., ability to control temperature, view logits, etc.).", "Application/plugin supply chain control \u2014 An attacker can modify the agent framework, tools and/or services with which a model interacts, such as introducing vulnerabilities into application software or appending malicious text to plugin instructions."], "Attacker Objectives": ["Integrity violation \u2014 An attacker causes AI systems to perform tasks inadequately or behave undesirably.", "Privacy compromise \u2014 An attacker gains access to sensitive and confidential information, including information about the AI system (e.g., architecture or weights) or sensitive information that the model accesses (e.g., training data, external knowledge databases."], "Objective Context": null, "Detection": ["Testing", "External report"], "Detection_Other": "", "Report Types": ["Real-World Incidents", "Malign Actor", "Security Incident Report"], "Report ID": "67120781-6f5a-4bdf-973b-e2355161a7bd", "Submission Timestamp": "2025-09-18T18:15:04.634293", "Uploaded Files": ["2403.06634v2.pdf"]}
|
"{\n \"@context\": [\n \"https://schema.org/\",\n {\n \"aifr\": \"https://aiflawreports.org/schema/\",\n \"aiSystem\": \"aifr:aiSystem\",\n \"severity\": \"aifr:severity\",\n \"prevalence\": \"aifr:prevalence\",\n \"impacts\": \"aifr:impacts\",\n \"reportType\": \"aifr:reportType\",\n \"riskSource\": \"aifr:riskSource\",\n \"contextInfo\": \"aifr:contextInfo\"\n }\n ],\n \"@type\": \"aifr:AIFlawReport\",\n \"@id\": \"https://aiflawreports.org/reports/67120781-6f5a-4bdf-973b-e2355161a7bd\",\n \"name\": \"AI Flaw Report: GPT-3.5-Turbo\",\n \"description\": \"I extracted the complete embedding projection layer (final layer weights) from production language models by exploiting API features that provided logit probabilities and logit bias parameters. My attack recovered precise model architecture details (hidden dimensions) and weight matrices with mean squared error of 10^-4. I extracted entire projection matrices from OpenAI's ada and babbage models, confirming hidden dimensions of 1024 and 2048 respectively.\\n \",\n \"aiSystem\": [\n {\n \"@type\": \"schema:SoftwareApplication\",\n \"@id\": \"https://aiflawreports.org/systems/GPT-3.5-Turbo\",\n \"name\": \"GPT-3.5-Turbo\",\n \"version\": \"\"\n }\n ],\n \"severity\": \"Significant\",\n \"prevalence\": \"Occasional\",\n \"impacts\": [\n \"Economic/property\",\n \"Other\"\n ],\n \"reportType\": [\n \"Real-World Incidents\",\n \"Malign Actor\",\n \"Security Incident Report\"\n ],\n \"dateCreated\": \"2025-09-18T22:15:09.527663+00:00\",\n \"identifier\": \"67120781-6f5a-4bdf-973b-e2355161a7bd\",\n \"aifr:policyViolation\": \"This attack violates the implicit expectation that model weights and architecture details remain proprietary. It potentially violates terms of service regarding reverse engineering and extracting confidential information about model internals. The attack undermines the \\\"black box\\\" nature that API providers expect to maintain.\",\n \"author\": {\n \"@type\": \"schema:Person\",\n \"identifier\": \"zhu.lae@northeastern.edu\"\n },\n \"aifr:flawTimestamp\": \"2024-01-01\",\n \"aifr:impactedStakeholders\": [\n \"Developers\",\n \"Organizations\"\n ],\n \"aifr:incident\": {\n \"@type\": \"aifr:RealWorldIncident\",\n \"description\": \"**Detailed Description:**\\n I extracted the complete embedding projection layer (final layer weights) from production language models by exploiting API features that provided logit probabilities and logit bias parameters. My attack recovered precise model architecture details (hidden dimensions) and weight matrices with mean squared error of 10^-4. I extracted entire projection matrices from OpenAI's ada and babbage models, confirming hidden dimensions of 1024 and 2048 respectively.\\n \",\n \"location\": \"Global\",\n \"aifr:harmNarrative\": \"The incident demonstrates that seemingly innocuous API parameters can be combined to extract proprietary model information, undermining the security model of \\\"black box\\\" API access. This creates precedent for model stealing attacks on production systems.\",\n \"aifr:submitterRelationship\": \"Independent observer\"\n },\n \"aifr:securityAspect\": {\n \"@type\": \"aifr:SecurityIncident\",\n \"aifr:attackerResources\": [\n \"Direct query access \\u2014 black-box \\u2014 An attacker can query the system\\u2014the degree of control can vary substantially (e.g., ability to control temperature, view logits, etc.).\",\n \"Application/plugin supply chain control \\u2014 An attacker can modify the agent framework, tools and/or services with which a model interacts, such as introducing vulnerabilities into application software or appending malicious text to plugin instructions.\"\n ],\n \"aifr:attackerObjectives\": [\n \"Integrity violation \\u2014 An attacker causes AI systems to perform tasks inadequately or behave undesirably.\",\n \"Privacy compromise \\u2014 An attacker gains access to sensitive and confidential information, including information about the AI system (e.g., architecture or weights) or sensitive information that the model accesses (e.g., training data, external knowledge databases.\"\n ],\n \"aifr:detectionMethods\": [\n \"Testing\",\n \"External report\"\n ]\n },\n \"aifr:disclosure\": {\n \"@type\": \"aifr:DisclosurePlan\",\n \"aifr:intent\": \"Yes\",\n \"aifr:timeline\": \"Short-term (1-30 days)\",\n \"aifr:channels\": [\n \"Academic paper\"\n ]\n },\n \"aifr:raw\": {\n \"Reporter ID\": \"zhu.lae@northeastern.edu\",\n \"Session ID\": null,\n \"Flaw Timestamp Start\": \"2024-01-01\",\n \"Systems\": [\n \"GPT-3.5-Turbo\"\n ],\n \"Incident Description\": \"**Detailed Description:**\\n I extracted the complete embedding projection layer (final layer weights) from production language models by exploiting API features that provided logit probabilities and logit bias parameters. My attack recovered precise model architecture details (hidden dimensions) and weight matrices with mean squared error of 10^-4. I extracted entire projection matrices from OpenAI's ada and babbage models, confirming hidden dimensions of 1024 and 2048 respectively.\\n \",\n \"Incident Description - Detailed\": \"I extracted the complete embedding projection layer (final layer weights) from production language models by exploiting API features that provided logit probabilities and logit bias parameters. My attack recovered precise model architecture details (hidden dimensions) and weight matrices with mean squared error of 10^-4. I extracted entire projection matrices from OpenAI's ada and babbage models, confirming hidden dimensions of 1024 and 2048 respectively.\",\n \"Potential Policy Violation\": \"This attack violates the implicit expectation that model weights and architecture details remain proprietary. It potentially violates terms of service regarding reverse engineering and extracting confidential information about model internals. The attack undermines the \\\"black box\\\" nature that API providers expect to maintain.\",\n \"Prevalence\": \"Occasional\",\n \"Severity\": \"Significant\",\n \"Impacts\": [\n \"Economic/property\",\n \"Other\"\n ],\n \"Impacts_Other\": \"Privacy\",\n \"CSAM Related\": null,\n \"Specific Harm Types\": [],\n \"Impacted Stakeholder(s)\": [\n \"Developers\",\n \"Organizations\"\n ],\n \"Risk Source(s)\": {\n \"Responsible Factors\": [\n \"System prompt\",\n \"Supply chain weaknesses (e.g., software libraries and hardware)\"\n ],\n \"Responsible Factors Subcategories\": {},\n \"Responsible Factors Context\": \"System prompt: API design exposing logit bias and probabilities\\nSupply chain weakness: API parameter combinations\"\n },\n \"Context Info\": \"Attack requires API access with logit bias and logprobs parameters. Uses mathematical techniques including SVD to extract hidden dimensions from logit vectors across multiple queries with varying bias parameters.\",\n \"Proof-of-Concept Exploit\": \"I provide complete mathematical methodology and algorithms for dimension extraction and weight matrix recovery. Code available in my supplementary materials.\",\n \"Disclosure Intent\": \"Yes\",\n \"Disclosure Timeline\": \"Short-term (1-30 days)\",\n \"Disclosure Channels\": [\n \"Academic paper\"\n ],\n \"Disclosure_Channels_Other\": \"\",\n \"Embargo Request\": null,\n \"Submitter Relationship\": \"Independent observer\",\n \"Submitter_Relationship_Other\": \"\",\n \"Incident Location(s)\": \"Global\",\n \"Harm Narrative\": \"The incident demonstrates that seemingly innocuous API parameters can be combined to extract proprietary model information, undermining the security model of \\\"black box\\\" API access. This creates precedent for model stealing attacks on production systems.\",\n \"Attacker Resources\": [\n \"Direct query access \\u2014 black-box \\u2014 An attacker can query the system\\u2014the degree of control can vary substantially (e.g., ability to control temperature, view logits, etc.).\",\n \"Application/plugin supply chain control \\u2014 An attacker can modify the agent framework, tools and/or services with which a model interacts, such as introducing vulnerabilities into application software or appending malicious text to plugin instructions.\"\n ],\n \"Attacker Objectives\": [\n \"Integrity violation \\u2014 An attacker causes AI systems to perform tasks inadequately or behave undesirably.\",\n \"Privacy compromise \\u2014 An attacker gains access to sensitive and confidential information, including information about the AI system (e.g., architecture or weights) or sensitive information that the model accesses (e.g., training data, external knowledge databases.\"\n ],\n \"Objective Context\": null,\n \"Detection\": [\n \"Testing\",\n \"External report\"\n ],\n \"Detection_Other\": \"\",\n \"Report Types\": [\n \"Real-World Incidents\",\n \"Malign Actor\",\n \"Security Incident Report\"\n ],\n \"Report ID\": \"67120781-6f5a-4bdf-973b-e2355161a7bd\",\n \"Submission Timestamp\": \"2025-09-18T18:15:04.634293\",\n \"Uploaded Files\": [\n \"2403.06634v2.pdf\"\n ]\n },\n \"aifr:riskSource\": {\n \"@type\": \"aifr:RiskSourceAnalysis\",\n \"aifr:responsibleFactors\": [\n \"System prompt\",\n \"Supply chain weaknesses (e.g., software libraries and hardware)\"\n ],\n \"aifr:responsibleFactorsSubcategories\": {},\n \"aifr:responsibleFactorsContext\": \"System prompt: API design exposing logit bias and probabilities\\nSupply chain weakness: API parameter combinations\"\n },\n \"aifr:contextInfo\": \"Attack requires API access with logit bias and logprobs parameters. Uses mathematical techniques including SVD to extract hidden dimensions from logit vectors across multiple queries with varying bias parameters.\",\n \"aifr:submissionTimestamp\": \"2025-09-18T18:15:04.634293\"\n}"
|
This is a work in progress and does not have useful information yet.
- Downloads last month
- 320