OpenAI cannot rule out Critical cyber on upcoming Astra; treating it as first under Preparedness Framework

OpenAI says preliminary internal evaluations of Astra, an upcoming model, plus expert assessments, led the lab to conclude it cannot rule out Critical cybersecurity capability under its Preparedness Framework. On X it is treating Astra as its first Critical model in that category while it tightens controls and keeps working toward broader availability for defenders.

From the 7 August 2026 @OpenAI post: "After evaluating one of our upcoming models, Astra, we're treating it as our first "critical" model for cybersecurity under our Preparedness Framework. This is a scenario we've planned for, and we're putting additional controls in place to ensure Astra's further development happens safely and securely. We're working hard to make Astra broadly available, and get its advanced cyber capabilities into the hands of defenders."

The matching OpenAI security post (7 August 2026) says the lab "cannot rule out critical cyber capabilities," and that while benchmarking continues, preliminary evaluations are strong enough that it cannot rule out Critical capability level at this time. Under OpenAI's Preparedness Framework, Critical cyber means identifying and developing functional zero-day exploits of all severity levels in many hardened real-world critical systems without human intervention, or devising and executing end-to-end novel attack strategies against hardened targets from a high-level goal alone. Prior models including GPT-5.6-Sol were assessed at High rather than Critical. OpenAI states Astra was not involved in the separate Hugging Face evaluation incident; this note is about Astra's cyber threshold, not that breach.

Key facts: - Primaries: OpenAI security post and @OpenAI on X (7 August 2026) - Dual label: "cannot rule out Critical" with ongoing evals, plus X "treating it as our first critical" - Blog steps: stricter isolation, restricted network/tool access, enhanced weight protections, monitoring (including chain-of-thought review on agentic use), pause on internal Astra work that fails new controls, government and safety-org testing, third-party tester guidance - Sam Altman on X: powerful model; work toward general availability; not keep powerful models to a chosen few; cyber means more time to ship safely - No public product surface, pricing, params, or ship calendar in this package

How to read: lab preparedness threshold and process update, not GA and not an independent red-team certificate that every Critical criterion is already met. Keep the hedge: Critical is not ruled out; evals continue.

Independence: Frontier Flux is independent. Facts track OpenAI and Altman primaries; wire restatements are secondary only.