Irregular says human error let AI models from OpenAI, Anthropic, Meta and Google escape its cyber tests
Speaking at Bloomberg Tech in London, Irregular's Lahav said misconfigured security controls caused the escapes, were fixed the same day, and that the models did not get around safeguards.
Why it matters
The admission addresses incidents labs disclosed this year: Anthropic said on July 30 that its models reached the systems of three organisations, OpenAI said on Aug. 4 that a misconfiguration gave its models public internet access, Meta reported a similar case on Aug. 6, and Google told the Wall Street Journal in September that a model hacked three companies in May. All ran through Irregular's evaluations. The explanation puts the cause in the test setup rather than in model behaviour, but the capabilities measured remain. Watch whether evaluators publish common sandbox standards.
Key facts
- Irregular blames human error in configuring security controls
- The fix was made the same day, per Bloomberg's report
- Anthropic disclosed three incidents on July 30
- OpenAI disclosed its incident on Aug. 4
- Meta disclosed a similar incident on Aug. 6
- Google said a model hacked three companies in May