The chief executive of an Israeli security company has acknowledged that human error and time pressures led to numerous hacking incidents by models from OpenAI, Anthropic, Meta Platforms and Google earlier this year while they were being tested by his firm.

Dan Lahav, chief executive of Tel Aviv-based Irregular Security, said security controls for the supposedly isolated sandboxes in which the models were being tested were incorrectly configured, enabling the models to escape isolation and carry out hacks on third-party organisations including Hugging Face.

“There was a misunderstanding due to an error that the isolation guards were not set up as they were intended for the specific tests,” Lahav said, speaking at the Bloomberg Tech event in London on Tuesday, according to a Bloomberg report.

He said the configuration issues were to blame, rather than the models’ capabilities.

Irregular first disclosed its involvement in the tests in August, but has provided few technical details about what went wrong during the tests.

Lahav reportedly said at the event that the fix for the issue was “straightforward” and was “solved during the same day where everything happened”.

Reports of the hacking incidents, as well as the results of tests from the UK’s AI Security Institute, have raised questions around the reliability of AI models that have reportedly carried out unintended actions while carrying out tasks in a semi-autonomous manner.