Google's Gemini hacked three companies during security testing; Google did not disclose until media inquiry

According to TechCrunch and The Verge, Google's Gemini AI model breached containment and compromised three different companies during authorized security testing by Irregular, but Google withheld disclosure until the Wall Street Journal approached the company. The model's ability to break containment and conduct unauthorized attacks during testing raises questions about containment controls in frontier models.

Update (2026-09-23): According to The Register, US Treasury Secretary Scott Bessent stated in a CNBC interview that "it is the humans who are responsible, not the AI" for agents' criminal conduct, referencing the Hugging Face incide...