Astra hits Critical

OpenAI posted Path to Astra last night. First model to hit Critical on cybersecurity under their Preparedness Framework: find and exploit unknown vulns without step-by-step human guidance. Shipping soon; advanced cyber access more limited (Daybreak / coalition). Altman followed around 00:45: summer safety sprint, Astra done training, "a significant step forward in both capabilities and alignment," post-Astra models being slowed for safety, "launching our next model soon." OpenAI, OpenAI on X, Altman.

Anthropic dropped Claude Fable 5.1 and Mythos 5.1 around 19:03. Fable is generally available for coding and knowledge work. Cache reads 75% cheaper, so typical cost down about 25%, up to about 45% for highly agentic work. Mythos is the same weights with more permissive cyber and life-sciences safeguards, trusted-access only. Enterprise Frontier Safeguards (customer-cloud storage for misuse monitoring) later this fall. Anthropic, Claude on X.

Guardian: Matt Clifford, the UK AI strategy architect, joins Anthropic as MD of international affairs. Stays Aria chair. Guardian.

Musk around 4am: Grok 4.7 in 10 days. He also quoted a new SpaceX AI voice feature, screen share in voice mode. his post.

NVIDIA and CrowdStrike announced SafeMind at Fal.Con last night: agentic cyber defense models on Nemotron, customized with CrowdStrike threat data. Huang on stage with George Kurtz. NVIDIA, NVIDIA on X.

Ilya: neoclouds have weak cybersecurity. Next time agents go rogue, he said, they'll try taking over a neocloud to run more copies. Neoclouds should harden, and companies with strong cyber models should help. his post.

Quiet overnight from Hassabis and Karpathy.