Lead: GPT-6 Astra hits OpenAI’s Critical cyber threshold
OpenAI’s GPT-6 Astra is the first model the lab classifies at the Critical level for cybersecurity under its Preparedness Framework. That bar means, with tools and access, the model can find previously unknown flaws and build novel exploit paths across hardened systems without continuous human guidance — including expert-supervised runs that produced new browser and kernel vulnerabilities and a working unsandboxed browser exploit chain.
OpenAI is pairing the designation with tighter monitoring, restricted cyber capability paths for general users, and a Daybreak Blue track aimed at trusted defenders. The same rollout broadens distribution through ChatGPT tiers, the API, and AWS, with Microsoft listing Astra in Azure AI Foundry the same day. For security and platform teams, the implication is concrete: treat Astra-class agents as high-capability offensive tooling by default, and design credentials, network scope, and human checkpoints accordingly.
What else matters
OpenAI ships Astra for Law
OpenAI launched Astra for Law — GPT-6 Astra plus a legal search index and practice-tuned instructions — for selected Am Law–oriented firms via a Trusted Access programme in ChatGPT and Codex, with API access coming for builders including Harvey and Legora. The package adds zero data retention on the API for eligible firms, 26 partner plugins (Thomson Reuters, iManage, Clio and others), and ChatGPT for Word GA. Legal tech vendors get a frontier model hook; firms get confidentiality controls before general availability.
Figure Helix 2.5: zero-shot across 30 homes
Figure’s Helix 2.5 completed long-horizon household tasks — tidying, folding towels, making beds — in 30 previously unseen Bay Area homes with no fine-tuning on those spaces. Index pretraining alone lifted full-task zero-shot success from 9% to 56%, and Figure reports a human-to-humanoid transfer scaling law as Index data doubles. The company says Index now adds ~35 minutes of human experience per second and has committed $3.5B of Helix compute — evidence that generalist home robots may transfer from human video before site-specific teaching.
Stanford’s 37,000-agent virtual biotech lands in Science
Stanford Medicine researchers published in Science a fully agentic “virtual biotech” of ~37,000 AI scientists spanning target discovery through trial design. Agents proposed a B7-H3 antibody-drug conjugate for lung cancer using only pre-January 2025 information; a major pharma later independently pursued the same strategy and secured FDA breakthrough therapy designation. James Zou’s group is now moving other AI-surfaced targets into wet-lab validation — a rare closed loop from agent design to independent clinical signal.
Worth your time
Odyssey-3: one world model, many bodies
Odyssey’s Odyssey-3 is an autoregressive diffusion world model trained to control robots, vehicles, drones, and even games from a shared physical prior with only hours of task-specific data. A useful read if you care where robotics and sim stack meet foundation models.
