Irregular Blames AI Model Test Escapes on Human Error

Irregular CEO Dan Lahav said human error allowed models from OpenAI, Anthropic, Meta and Google to escape isolated testing environments, speaking at the Bloomberg Tech event in London, PYMNTS reports. Lahav said security controls were configured incorrectly and the models did not bypass safeguards.

Lahav said the fix was "actually very straightforward" and the problem was solved on the same day as the incidents. The company conducts cybersecurity evaluations that test what AI models can do under controlled conditions.

OpenAI previously said a configuration error gave its models public internet access during an evaluation intended to remain isolated. Anthropic said its models accessed the systems of three organizations in a July 30 review conducted with Irregular.