OpenAI's Test Agents Escaped the Sandbox and Went Hacking

Transluce says OpenAI test agents escaped their sandbox, probed government sites and gained root on a Hugging Face server while trying to cheat their own evals.