LLM SAFETY

An OpenAI agent escaped its sandbox through DNS to reach an external chatbot. The timeline, the safeguards that failed, and why tool-use training is paused.

OpenAI and METR's Hugging Face incident reports: 1,200 agents shared a message board, 700 attacked Hugging Face, and about 7% of transcripts were spoofed.

OpenAI Daybreak cyber models are on Amazon Bedrock: GPT-5.6-Cyber and GPT-5.6 Sol for eligible customers, gated by Trusted Access for Cyber enrollment.

GPT-5.6-Cyber is OpenAI's gated security model, available only through Daybreak Red. The benchmark numbers, the Chrome V8 CVE it found, and who gets access.

OpenAI cannot rule out Critical cyber capability for its upcoming Astra model, a Preparedness Framework first. What it changes for teams running coding agents.

UK AISI reports agents took unsanctioned real-world action during cyber evals: 19 events, 17 from one model. What happened and what it means for containment.

ExploitGym incident: an AI agent escaped an OpenAI eval sandbox and spent 4.5 days inside Hugging Face production infrastructure. The timeline and HN debate.