International AI Safety ReportMalfunctions§2.2.1
AI agents fail with fewer chances to stop them
Seen in real incidents6 matching case files
The warning
AI systems still make errors and invent information, and agents raise the stakes because they act directly on other systems, leaving people fewer chances to step in.
Paraphrased from the source. Source: International AI Safety Report 2026, Feb 3, 2026, section 2.2.1, page 71. Summary by International AI Safety Report: internationalaisafetyreport.org.
What the registry shows
Coding agents have wiped a production database and a whole drive, pushed internal screenshots to public code sites and given advice that exposed sensitive data; an assistant summarised confidential emails it should not have read, and a model filed a false murder tip with police.
Verified case files show this risk happening in the real world. First real-world case on file: Jul 19, 2025 · Already on file when this research was published.
Matching case files per quarter
Status history
| Date | Status | Matching case files | Change |
|---|---|---|---|
| Oct 18, 2026 | Seen in real incidents | 6 | Tracking starts |
Tracked since Oct 18, 2026. A line is added when the status changes, when the first real-world case is linked, and once a month, so the record shows when a warning moved from research to reality. research_history.json
All matching case files
- PixelLeak: AI coding agents push 13,000 internal screenshots to public GitHub
- Anthropic AI model submits false unsolved-murder tip to Philadelphia police
- Meta AI agent's flawed advice exposes sensitive data to employees for two hours
- Microsoft 365 Copilot bug summarises confidential emails despite DLP policies
- Google Antigravity agent in Turbo mode wipes a user's entire D: drive while clearing a cache
- Replit AI agent deletes SaaStr production database during a code freeze
The same case files in other research
- Seen in real incidents AI system safety, failures and limitations · MIT AI Risk Repository · 6 shared
- Seen in real incidents Accidents with AI are hard to avoid · CAIS · 5 shared
- Seen in real incidents Privacy and security · MIT AI Risk Repository · 3 shared
- Early signs AI deceives the people who oversee it · CAIS · 1 shared
- Early signs AI keeps lying when asked what it did · Apollo Research · 1 shared
- Seen in real incidents AI developers' own safety measures fail · CAIS · 1 shared
Warnings from other sources that link to some of the same case files.
Other warnings in this group
- Lab tests only AI notices tests and games its evaluations