OpenAI says its forthcoming Astra model is the first system to meet the Critical cybersecurity capability bar in the company’s Preparedness Framework. In plain terms, OpenAI judges that with the right tools and access, Astra can find previously unknown security flaws and develop ways to exploit them across well-protected systems without a person guiding each step.
The lab delayed parts of Astra’s development and release while it hardened protections against cyber misuse and unauthorized model actions, then concluded safeguards are strong enough to move toward release. OpenAI stresses Astra was not involved in the earlier Hugging Face agent breakout, but says learnings from that incident shaped the new controls. A public launch is described as “soon,” with the fullest cyber powers gated rather than handed to every ChatGPT user on day one.
Source: https://openai.com/index/path-to-astra/