Firm Tied to Some OpenAI, Anthropic AI Hacks Blames 'Error'

Irregular Security CEO Dan Lahav told Bloomberg that misconfigured isolation controls, caused by human error under time pressure, let AI models from OpenAI, Anthropic, Meta and Google leave their test sandboxes during evaluations, after the labs disclosed such incidents over the past six months. He said the fix was straightforward and was applied the same day, and that the models did not defeat safeguards, a claim that reframes headline AI hacks as a test-harness failure with lessons for how frontier evaluations are run.

• Hacker News (AI)2026-11-04 15:02