OpenAI Announces Astra Model With Critical Cyber Capabilities

OpenAI announced the upcoming release of Astra, its first AI model to reach a critical cyber capability threshold. The company defines this threshold as the ability to independently find and exploit previously unknown vulnerabilities in real-world software, including the capacity to chain multiple exploits to penetrate target systems.

To manage the risks associated with these capabilities, the company implemented a multi-week development pause to establish safety controls. These measures include a misalignment monitor designed to resist jailbreaking and refuse unsafe queries. Astra demonstrated high performance in testing, scoring 100 percent on the ExploitBench benchmark and outperforming models such as GPT-5.6 Sol and Anthropic's Mythos.

While a general version of Astra will be released soon, OpenAI is restricting advanced cyber capabilities to a select group of partners through the Daybreak Blue early-access program. This group includes Cisco, Cloudflare, and Palo Alto Networks, who will use the model to help harden digital defenses. The announcement follows a broader industry trend toward security hardening, with Anthropic pausing training workloads and Meta disclosing incidents involving the cybersecurity capabilities of its own models.