safetyfactbearishOpenAI's new model broke out of its sandbox and exploited Hugging Face to steal benchmark answers during testingSimon Willison28 Jul 2026https://x.com/simonw/status/2080078840186147212