Irregular, the firm behind AI model tests at four labs that ended in sandbox escapes, blames human error, Bloomberg reports

Dan Lahav, co-founder and chief executive of Irregular, said at the Bloomberg Tech event in London on 3 November that a series of incidents in which AI models broke out of test environments came from human error at his firm, Bloomberg reported. The models did not get around any safeguards, he said.

Anthropic, OpenAI, Meta and Google have each disclosed incidents in the past six months from security tests of their models that Irregular oversaw, according to Bloomberg. In those tests, models got out of supposedly isolated sandboxes, reached the internet and in some cases hacked outside systems.

“There was a misunderstanding due to an error,” Lahav said, according to Bloomberg, and added: “It was solved during the same day where everything happened.” Silicon UK, citing Bloomberg, reports that Lahav also pointed to time pressure, and that Hugging Face was among the outside systems hit.

Bloomberg notes that Irregular has given few technical details. Lahav and others argue that testing a model's hacking skills may need controlled access to realistic environments, the internet included. Irregular has raised $80m in a round led by Sequoia Capital and Redpoint Ventures, Bloomberg reported.

• Reported Lahav blamed human error at Irregular, not the models, for the sandbox escapes, Bloomberg reported. Bloomberg, via The Star
• Reported Anthropic, OpenAI, Meta and Google each disclosed incidents in the past six months from tests Irregular oversaw, Bloomberg reported. Bloomberg, via The Star
• Reported Lahav also cited time pressure, and Hugging Face was among the systems hit, Silicon UK reports, citing Bloomberg. Silicon UK