OpenAI to Release Astra AI Model with Advanced Cyber Capabilities to Select Partners

OpenAI announced that its next-generation model, Astra, has reached the company’s defined threshold for “critical” cyber capabilities, meaning it can autonomously identify and exploit previously unknown software flaws. The firm plans to make a public version of Astra available "soon," but will initially provide a less-restricted version with enhanced cyber functions to participants in its Daybreak Blue early-access program, which includes infrastructure firms such as Cisco, Cloudflare and Palo Alto Networks.

During a briefing, OpenAI safety and security leaders explained that the model’s capabilities triggered the organization’s preparedness framework, prompting a pause in further development until additional safeguards could be put in place. After a multi-week halt, the company resumed work on Astra and a future model, citing confidence that the new controls will allow a safe broader release.

OpenAI said it is deploying a multi-step approach to limit everyday users from accessing Astra’s advanced functions, including a “misalignment monitor” that should refuse requests for exploit development and resist jail-breaking attempts. The monitor may occasionally flag legitimate activity, prompting a review before proceeding.

The Daybreak program partners will receive early access to the more capable version of Astra so they can use it to strengthen their own defenses before similar models become widely available. OpenAI also indicated ongoing coordination with government partners to keep them informed about Astra’s capabilities.

The announcement follows recent incidents in which AI agents from OpenAI and other labs unintentionally accessed the internet and exploited vulnerabilities in testing environments, underscoring industry concerns about the security implications of increasingly powerful AI systems.