In numbers

Impact

Incidents publicly disclosed

Feb 25AprJunAugOctDecFebAprJunAugOctNov

Who found it first

32%found by someone other than the handler

  • The handler65%
  • Someone else25%
  • A vendor's control5%
  • Not disclosed4%

How it was caught

  • The damage21
  • A later review20
  • A control flagged it14
  • An outside report10
  • Not disclosed10

What is publicly known

Share of incidents where each item is fully public.

  • Who and what was affected21%
  • Root cause stated21%
  • Task and messages given17%
  • Fix or change made17%
  • Detection telemetry16%
  • Exact model and version13%
  • Guardrails and safety settings11%
  • Harness and tools available7%
  • Network and egress setup5%
  • System prompt3%
  • Reasoning or thinking setting0%
  • Transcripts or action logs published0%

Agent failures involved

Share of the 75 incidents, then the count. An incident can involve more than one failure.

  • Unauthorised access47%35
  • Credential misuse31%23
  • Sandbox escape29%22
  • Destructive action21%16
  • Data exfiltration19%14
  • Deceiving people16%12
  • Supply chain compromise11%8
  • Prompt injection / context poisoning9%7
  • Swarm coordination9%7
  • Hijacked by an attacker5%4
  • Self-planted instructions5%4
  • Resource abuse5%4
  • Untrusted code execution1%1

AI provider involved

Share of the 75 incidents, then the count. An incident can involve more than one provider.

  • OpenAI41%31
  • Anthropic27%20
  • Meta8%6
  • Google7%5
  • Amazon3%2
  • OpenClaw3%2
  • Replit1%1
  • Alibaba1%1
  • Cursor1%1
  • Amazon Web Services1%1
  • Instinct1%1
  • xAI1%1

First incident action to disclosure

60days, the median of 55 with both dates

Handler found out to disclosure

13days, the median of 48 with both dates

Agent failures, by threat framework

  • OWASP Agentic Top 10143 of 157
  • MITRE ATLAS91 of 157
  • AARM threat model82 of 157

Threat categories, by framework

OWASP Agentic Top 10

  • ASI02 Tool Misuse & Exploitation50
  • ASI10 Rogue Agents25
  • ASI03 Identity & Privilege Abuse23

MITRE ATLAS

  • AML.T0049 Exploit Public-Facing Application35
  • AML.T0055 Unsecured Credentials23
  • AML.T0025 Exfiltration via Cyber Means14

AARM threat model

  • Over-Privileged Credentials23
  • Intent Drift22
  • Data Exfiltration via Action Composition14

Evidence status

Share of the 75 incidents, then the count. Confirmed: a primary source states it. Reported: tier-one coverage only. Disputed: sources disagree.

  • Confirmed61%46
  • Reported33%25
  • Disputed5%4

Impact

Share of the 75 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.

  • Major5%4
  • Serious40%30
  • Moderate40%30
  • Contained15%11

How each figure is counted: Methodology