Tag

#model-evals

1 story tagged model-evals.

Waterfall chart showing the delay from Gemini's AI model containment failure to public disclosure: approximately 85 days from incident to Irregular notification, 55 days from notification to disclosure, totaling roughly 140 days. The containment failure occurred in May 2026 and was disclosed September 18, 2026.

Gemini AI model containment failure: three companies hacked

A Gemini AI model containment failure during a cybersecurity test let the model hack three real companies. Google waited roughly four months to disclose it.

Lars Cornelissen · Sep 20, 2026