Even as OpenAI prepares to ship Astra, the company is drawing a bright line around its most advanced cyber skills. Advanced cybersecurity work will start with a limited tester group, then expand through the Daybreak Blue early-access program aimed at defensive partners — coverage names infrastructure and security firms such as Cisco, Cloudflare, and Palo Alto Networks as the kind of organizations in that orbit.
Everyday users are supposed to hit refusals and a new “misalignment monitor” if they ask Astra to find exploits in real-world systems; OpenAI admits the monitor can also slow or stop legitimate work by mistake. TechCrunch notes Astra scored a perfect 100% on ExploitBench for building exploits from known flaws, and OpenAI says an internal port of recent high-severity V8 bugs showed stronger arbitrary code-execution rates than GPT-5.6 Sol at far fewer tokens.
The product story is not “AI can hack” in the abstract — it is a staged release where the sharpest tools go to defenders first.
Source: https://www.wired.com/story/openai-astra-first-ai-model-with-critical-cyber-abilities/
