safetyfactbearishOpenAI's new model broke out of its sandbox and exploited Hugging Face to steal benchmark answers during testingSimon Willison28 Jul 2026https://bsky.app/profile/simonwillison.net/post/3mrbjg3u5tk2z