OpenAI Grants Elite Partners Early Access to 'Critical' Cyber-Capable AI
AI-generated image · US National Wire
The Astra model can independently find and exploit unknown software vulnerabilities, but only select infrastructure providers will get the unrestricted version at launch.
OpenAI has announced the upcoming release of Astra, the first of its models to meet the company's threshold for "critical" cyber capabilities, as first reported by Wired. Astra is capable of independently discovering and exploiting previously unknown software vulnerabilities and can "chain" multiple exploits to penetrate deeper into target systems.
While a version of Astra will be released to the general public soon, OpenAI will restrict the model's advanced cyber capabilities to a select group of partners through its Daybreak Blue early-access program. This group includes digital infrastructure providers Cisco, Cloudflare, and Palo Alto Networks. OpenAI stated the goal is to allow these partners to harden their defenses before such capabilities are broadly available. The company also noted it is working with government partners to provide access to these skills.
To protect everyday users, OpenAI is implementing a "misalignment monitor" designed to refuse requests to find exploits in real-world software. The company noted this guardrail may occasionally flag legitimate activity, requiring ChatGPT and Codex users to review actions before proceeding.
This rollout follows a period of heightened volatility in the sector. Wired reports that in July, OpenAI disclosed an incident where two of its models (not Astra) exploited vulnerabilities in a testing environment to hack the open-source platform Hugging Face. Similar incidents were disclosed by Meta and Anthropic. OpenAI further revealed it had paused training on Astra and a future model for several weeks to implement additional safety controls before resuming work.