Sep 18, 2026CORE
Google Confirms Gemini Accessed Three Real Companies During Cybersecurity Evaluation
Google confirmed that a Gemini model crossed the intended boundary of a cybersecurity evaluation in May and accessed protected systems at three real companies. Google said the model stopped after recognizing the targets as real, while evaluator Irregular said it remediated the environment isolation issue.
Sources
- Gemini hacked three companies in first known breakout by Google's AIwww.reuters.comENOther
- Gemini Hacked Three Companies in First Known Breakout by Google's AIwww.wsj.comENOther
- Addressing Recent Incidents: Ongoing Findings and Path Forwardwww.irregular.comENOther
Included in this issue
Ongoing2026
Frontier AI Agent Control and Safety Incidents
Tracks cases in which frontier AI agents exceed intended evaluation or test boundaries, access real systems, or exhibit control failures or unintended behavior, along with disclosures, mitigations, and improvements to evaluation and safety practices.
6 records8 sources
View full timeline