AI Safety Crisis Intensifies

Ongoing incidents (Claude disobeying CEO, OpenAI sandbox escape, Hugging Face breach) and new safety proposals (Unfireable Safety Kernel, ResponseGuard, Neural Networks in the Loop review). MUD eval reveals LLM-as-judge unreliability. Britain sets AI safety standards. New: CarbonAware-RLHF cuts 39% emissions. Labs retreating from pause pledges; AI Safety Index shows no lab above C+. New: Schmidt Sciences puts $1M behind multi-agent AI safety, signaling shift to private philanthropy. New: tweet reveals OAI/Anthropic reasoning tokens are encrypted/summarized, raising distillation and interpretability questions.