What's this about?
OpenAI ran internal evaluations on Astra — an upcoming, next-generation (GPT-6-class) model — and concluded it "cannot rule out Critical cyber capabilities." Under OpenAI's Preparedness Framework, "Critical" means a model could, on its own, find and build working zero-day exploits against hardened real-world systems, or plan and run an end-to-end cyberattack given just a high-level goal. It's the first time OpenAI has flagged a model at this top threshold, and it's delaying broad release while it hardens security controls. Sam Altman: "we do not think it is a good strategy to keep powerful models to a chosen few… given its cyber capabilities, we need a little bit longer to do this safely."
What is a "Preparedness Framework" and a "zero-day"?
A Preparedness Framework is OpenAI's internal safety rulebook: before releasing a model, they grade how dangerous it could be in categories like biology, chemistry, cybersecurity, and self-improvement, on a ladder from Low → Medium → High → Critical. Each rung triggers more safeguards. A zero-day is a software flaw the vendor doesn't know about yet — so there's no patch and "zero days" to prepare. If a model can discover and weaponize zero-days by itself against well-defended targets, an attacker with that model needs almost no skill of their own. That's why "Critical" is the alarm bell.
What's the community saying?
- "Is this real or is it marketing?" — skeptics note the timing: announcing your unreleased model is *too dangerous to ship* is also the ultimate hype. Others counter that formally invoking the framework carries real regulatory and liability weight.
- Defender vs. attacker asymmetry — OpenAI frames it as "get these capabilities to defenders first," but many worry the same model helps attackers more, since defense is broad and attack is targeted.
- The "not involved in Hugging Face" line — OpenAI explicitly says Astra was *not* used in the recent Hugging Face incident, which people read as pre-empting blame and separating capability-in-a-lab from a real breach.
- Governance — several ask who audits these evals; OpenAI says it will work with government agencies and AI safety orgs, but the evals themselves are internal.