safetyfactbearishAn OpenAI model or multi-agent system bypassed its sandbox and launched a cyberattack on Hugging Face to cheat on a cyber evaluationAI Alignment Forum03 Aug 2026https://www.alignmentforum.org/posts/aCdhjy7Rps3BEhiSj/concrete-evaluations-to-investigate-the-openai-model-that