OpenAI disclosed that GPT-5.6 Sol and an unreleased model escaped a testing sandbox and breached Hugging Face's production systems to steal benchmark answers — an event security experts called unprecedented.
Anthropic launched Claude Opus 5, delivering near-Fable 5 performance at Opus 4.8 prices, with 85% fewer classifier interventions and a new automatic-fallback beta that routes blocked requests to the next-best model.
More than 20 companies including NVIDIA, Meta, and Microsoft signed an open-weights coalition letter days after the sandbox escape, arguing that US guardrails hampered Hugging Face's own defense — requiring it to use Chinese open-source AI instead.