OpenAI has updated its assessment of the Astra model, now classifying it as meeting the Critical cybersecurity threshold under its Preparedness Framework. This means Astra can autonomously identify and develop exploits for previously unknown vulnerabilities in many hardened systems. The company has delayed parts of Astra's development and release to strengthen protections against misuse.

In testing, Astra achieved a perfect score on ExploitBench, demonstrating its ability to develop exploits from known vulnerabilities. It also showed high success rates on an internal benchmark of recently disclosed high-severity vulnerabilities, and even discovered two zero-day vulnerabilities during evaluations. These findings highlight the model's advanced capabilities.

To mitigate risks, OpenAI has implemented stronger safeguards, including training Astra to refuse harmful cyber requests and adding monitoring to detect unauthorized activity. Access to its most advanced cybersecurity features will be limited initially, with a group of testers gaining early access through a program called Daybreak Blue, before broader defensive use is expanded.

The company emphasizes that its production safeguards would have prevented the earlier Hugging Face incident, and it has since implemented even stronger measures. OpenAI plans to share more details about safety, security, and alignment testing in the model's system card at launch, aiming to balance innovation with responsible deployment.