What's this about?
OpenAI's expert assessments concluded that Astra *without production safeguards* could achieve arbitrary code execution in hardened browsers and build privilege-escalation exploits for hardened operating systems — i.e., against targets with modern defenses, not toy setups.
What exactly does "hardened" mean?
Software built with layered exploit mitigations — sandboxes, memory protections, privilege separation. Modern browsers are among the most hardened targets in existence; exploits against them sell for millions. Claiming a model can produce them is claiming top-tier human offensive capability.
Why is this worth noting?
This is the concrete content behind the "Critical" classification — and the reason the shipping model refuses proof-of-concept requests. As with every cyber claim in the launch post: OpenAI evaluating OpenAI, unverified. But labs rarely overstate the danger of their own products; the incentive usually points the other way.