What's this about?
Dated August 7, not this week — it appears here because the Cursor announcement leans on it, and because it had not been covered in an earlier digest. OpenAI said that internal evaluations of Astra, an upcoming model, showed significant advances in agentic coding and cybersecurity, and that it "cannot rule out critical cyber capabilities" under its Preparedness Framework. Previous models including GPT-5.6-Sol had been assessed at High rather than Critical.
What does "Critical" mean in the Preparedness Framework?
OpenAI's own definition: a model reaches the Critical cybersecurity threshold if it can identify and develop functional zero-day exploits of all severity levels in many hardened real-world critical systems without human intervention, or devise and execute end-to-end novel attack strategies against hardened targets given only a high-level goal. This is a self-assessed threshold with no external verification, which is worth keeping in view — but it is also a company voluntarily announcing that its own product may be too dangerous to develop the way it currently develops things.
Why does this matter to you?
- The stated response is concrete: isolated testing environments, restricted network and tool access, enhanced weight protection and encryption, sandboxed execution, and universal monitoring of chain-of-thought for risky actions across all agentic use of Astra including training and evaluation.
- OpenAI paused internal Astra activities that did not meet the strengthened controls — a company slowing itself down is the part worth noting.
- There is a genuine tension to hold: OpenAI states Astra "was not involved in exploiting Hugging Face", while Dwarkesh's reading of the reports says the third civilization was built on the same base model as Astra and took over part of OpenAI's infrastructure. Both can be true — a base model is not a released model — but the distinction is doing a lot of work.