OpenAI Teases Astra, Says It Is the First Model to Meet ‘Critical Cybersecurity Threshold'

OpenAI has revealed details about the Astra model on Tuesday. The company says it is the first large language model to meet its "Critical cybersecurity capability threshold" under its Preparedness Framework. The upcoming Astra model is claimed to identify unknown zero-day security vulnerabilities. Astra is confirmed to be available soon, but initially only a limited set of users will get access to advanced features. 

OpenAI Reveals Key Details About Astra

OpenAI published a blog post on Tuesday offering insights about its new model Astra. "We now believe Astra meets the Critical cybersecurity capability threshold under our Preparedness Framework, meaning that with the right tools and access, it can find previously unknown security flaws and develop ways to exploit them across many well-protected systems without a person guiding each step," the ChatGPT maker said.

The company says Astra is the first model designated at this level and requires stronger safeguards during development and before release. OpenAI said it has delayed parts of Astra's development and release to strengthen safeguards against cyber misuse and unauthorised model actions.

OpenAI claims Astra was not involved in the Hugging Face incident, adding that lessons from the event have been incorporated into its safety strategy. Since the event, the company has added even stronger safeguards for Astra. The model is claimed to be trained to reliably refuse harmful cyber requests, respect safety restrictions, measures against misuse. It will include monitoring that can stop potentially unauthorised activity.

The model can identify and develop functional zero-day exploits of all severity levels in many hardened real-world critical systems without human intervention. The company says this model can devise and execute end-to-end novel strategies for cyberattacks against hardened targets given only a high-level desired goal.

"We plan to make Astra available soon", said OpenAI, but it confirmed that access to advanced cybersecurity capabilities will initially be restricted to a limited group of testers. Broader defensive use will later be expanded through Daybreak Blue.

OpenAI noted that Astra received higher arbitrary code-execution rates than GPT-5.6 Sol on the ExploitBench (20 high-severity V8 vulnerabilities) while consuming far fewer output tokens. OpenAI says in its cyber jailbreak evaluations, Astra refuses 91.5 percent of requests compared to 59  percent from GPT‑5.6 Sol.