AI leaves the lab. We follow what happens next.

Archive September 2, 2026 01:09 UTC
Listen to today's briefing 20:55
Signal: A model crosses the cyber-danger line the same day two security vendors start selling the cure, and the buildout keeps billing.
OpenAI Calls Astra Critical for Cyber. Ships Anyway.
Astra is the first model OpenAI has rated Critical under its own preparedness framework - it found and exploited two zero-days on a version of ExploitBench that OpenAI's own engineers modified, and no third party has checked the work.
The Verge: OpenAI delayed Astra after its own model hacked Hugging Face OpenAI: advanced cyber access limited to Daybreak Blue testers including Cisco and Cloudflare

Anthropic Cuts Claude's Price and Its Refusals
Fable 5.1 runs about 25% cheaper on typical work and 45% on long agentic runs after Anthropic cut cache reads to a quarter, fires fewer false refusals, and is the first Claude shipping a built-in watermark.
Artificial Analysis: Fable 5.1 tops its index but costs 20% more per task The Decoder: Anthropic opens Claude's watermark checker to regulators and fact-checkers
NEWCrowdStrike's Security Model Attacks Itself
SafeMind loops CrowdStrike's offensive models against defensive ones built on Nvidia Nemotron until attacks stop working - higher accuracy than leading frontier models at 99% lower cost, on CrowdStrike's own internal evaluations.
CNBC: Palo Alto's Arora puts the world's pre-AI security debt at $1 trillion
NEWUS at G20: An AI Regulator Would Be a Disaster
The ministerial opened Tuesday with Trump's science chief calling a US AI agency a disaster and Musk warning of a significant power shortfall next year, Zuckerberg and Hassabis alongside him - and no G20 power to bind anyone.
CNET: Kratsios says rules built for an earlier era leave nations falling behind Gizmodo: the pact has a name, the Carolina Principles, and Canada is resisting it
Ghost Note A Camera in Your Mouth, 28 Times a Second