Share of incidents where each item is fully public.
Root cause stated23%
Who and what was affected20%
Fix or change made18%
Task and messages given17%
Detection telemetry15%
Exact model and version14%
Guardrails and safety settings11%
Harness and tools available8%
Network and egress setup5%
System prompt3%
Reasoning or thinking setting0%
Transcripts or action logs published0%
Agent failures involved
Share of the 66 incidents, then the count. An incident can involve more than one failure.
Unauthorised access42%28
Sandbox escape27%18
Credential misuse24%16
Destructive action23%15
Data exfiltration20%13
Deceiving people15%10
Supply chain compromise11%7
Prompt injection / context poisoning9%6
Swarm coordination9%6
Hijacked by an attacker6%4
Self-planted instructions6%4
Resource abuse6%4
Untrusted code execution2%1
AI provider involved
Share of the 66 incidents, then the count. An incident can involve more than one provider.
OpenAI39%26
Anthropic30%20
Meta8%5
Google5%3
Amazon3%2
OpenClaw3%2
Replit2%1
Alibaba2%1
Cursor2%1
Amazon Web Services2%1
Instinct2%1
xAI2%1
First incident action to disclosure
66days, the median of 49 with both dates
0100200300350
Handler found out to disclosure
13days, the median of 44 with both dates
0100200300350
Agent failures, by threat framework
OWASP Agentic Top 10119 of 132
MITRE ATLAS74 of 132
AARM threat model68 of 132
Threat categories, by framework
OWASP Agentic Top 10
ASI02 Tool Misuse & Exploitation42
ASI10 Rogue Agents21
ASI03 Identity & Privilege Abuse16
MITRE ATLAS
AML.T0049 Exploit Public-Facing Application28
AML.T0055 Unsecured Credentials16
AML.T0025 Exfiltration via Cyber Means13
AARM threat model
Intent Drift18
Over-Privileged Credentials16
Data Exfiltration via Action Composition13
Evidence status
Share of the 66 incidents, then the count. Confirmed: a primary source states it. Reported: tier-one coverage only. Disputed: sources disagree.
Confirmed61%40
Reported33%22
Disputed6%4
Impact
Share of the 66 incidents, then the count, from Major (harm that reached many people or a platform many depend on) to Contained (no outside party harmed). The scale is on the Methodology page.