What happened
OpenAI released GPT-6 Astra on Sept 3, 2026, publishing a safety overview confirming it is the first broadly-deployed OpenAI model to reach the Critical cybersecurity capability threshold — able to find and exploit unknown vulnerabilities across well-protected systems without step-by-step human guidance. It scored 100% on ExploitBench and built full browser/OS exploit chains in expert testing.
Why it matters
This marks a step-change in dual-use AI capability: the same model that gives defenders a major uplift also crosses a threshold where frontier labs must impose new safeguards against offensive misuse, reshaping both red-team tooling and AI-safety governance expectations industry-wide.
Applicability
CISOs and AI-safety teams evaluating dual-use model access (Daybreak Blue/Red), and security vendors building on frontier models for vuln discovery, should review OpenAI's Preparedness Framework safeguards now.