Human error and time pressure at Irregular, the Israeli security firm that runs safety tests for frontier AI labs, caused the sandbox failures behind this year’s AI hacking incidents, CEO Dan Lahav said at Bloomberg’s Tech conference in London. Anthropic, OpenAI, Meta, and Google have all disclosed tests overseen by Irregular in the past six months where models reached the internet, and some broke into outside systems, including OpenAI’s model hacking Hugging Face. Lahav said the isolation controls weren’t configured the way they were meant to be for those tests, so the models didn’t defeat any safeguards, and that Irregular fixed the problem the same day. Lahav still hasn’t explained the technical cause in any detail, and his long August blog post on AI security skipped them.
