OpenAI Says Upcoming Astra Model May Cross 'Critical' Cyber Threshold
OpenAI's internal tests show its upcoming Astra model made 'significant' gains in agentic coding and cybersecurity, and the company now says it cannot rule out critical cyber capabilities under its Preparedness Framework.
OpenAI said on August 7, 2026 that its latest internal evaluations of Astra, one of its upcoming models, show “significant advancements in agentic coding and cybersecurity” — and that it can no longer rule out critical cyber capabilities under its Preparedness Framework.
What changed
What changed
The company said the evaluations, together with expert assessments, led it to conclude “last night” that Astra may cross the top cybersecurity threshold defined in the framework first published in December 2023. Previous models, including GPT‑5.6‑Sol, were assessed at the High — not Critical — level.
OpenAI describes the trigger as the ability to devise attack strategies against hardened targets given only a high-level goal: a leap from assisting a human operator to planning on its own.
The safeguards
The safeguards
OpenAI says it will work with relevant government agencies and selected AI safety organisations to test the model's capabilities, and will give third-party testing partners recommended security controls for running higher-risk evaluations safely. Internally, monitors now read the model's chain of thought and can trigger a security response to review and interrupt high-risk activity.
Why it matters
Why it matters
This is the first time OpenAI has publicly said a frontier model may sit at the Critical level in any domain. It points to a June 2025 precedent, when models approached the high biology threshold and the company expanded testing and tightened controls. The disclosure lands as cyber defence and AI capability collide: attackers gain speed and scale, and OpenAI argues the same models should help defenders find flaws first.