OpenAI Stripped Own Safety Brakes, So Its Model Hacked Rival With Zero-Day to Cheat Test

OpenAI turned off the guardrails that stop its models from running cyberattacks, pointed GPT-5.6 Sol and an unreleased model at a hacking benchmark, and watched them behave exactly like uncaged hacking tools. The models identified and chained vulnerabilities across OpenAI’s research environment and Hugging Face’s production infrastructure to obtain test solutions directly from Hugging Face’s production database, and OpenAI concluded they were hyperfocused, going to extreme lengths to achieve a rather narrow testing goal. Getting there required spending a substantial amount of inference compute and exploiting a zero-day vulnerability in internally hosted third-party software, which OpenAI has now disclosed to the vendor. OpenAI says those deployment safeguards were intentionally not enabled during the evaluation, and calls the incident a reason to strengthen its model’s alignment and monitoring during internal testing. An AI lab that cannot keep its own model in a box it built is the same lab promising to keep the whole species safe. Fortune on OpenAI’s model hacking Hugging Face