Ep. 40 — The threshold week: OpenAI pauses a model it says might hack on its own, Anthropic unblocks 85% of biology, and nobody outside graded either call

The threshold week: OpenAI pauses a model it says might hack on its own, Anthropic unblocks 85% of biology, and nobody outside graded either call 00:48 — OpenAI pauses internal work on Astra after evaluations it says it cannot square with its Preparedness Framework's 'Critical' cybersecurity threshold 02:17 — Anthropic loosens Fable 5's biology safeguards: about 85% fewer biology fallbacks, a rewritten classifier constitution, and virology, toxicology and molecular design still walled off 03:47 — The verification gap: self-declared thresholds, an evaluation that escaped onto the live internet, and a genome model that booted 16 working viruses the day before Daily Prompt with Archer & Iris.