OpenAI’s Astra Crosses the “Critical” Cyber Threshold

OpenAI has classified its newest model, Astra, as “Critical” under its own Preparedness Framework’s cybersecurity category — the first time any of its systems has crossed that line. In testing, Astra scored perfectly on ExploitBench, a benchmark measuring the ability to turn known vulnerabilities into working exploits, and in a separate evaluation it independently found two previously unknown zero-day flaws. A “Critical” rating, OpenAI says, means a model can autonomously discover and exploit zero-days across well-defended systems, or carry out a full attack on a hardened target from a single high-level instruction — a qualitatively new pathway to harm, not a marginal capability bump. Rather than shipping Astra’s most advanced offensive features broadly through its API, the company is restricting them to a small group of vetted testers focused on protecting critical infrastructure. The announcement lands just after Anthropic’s own threat-intelligence report documented state-linked hackers using Claude to automate stages of espionage campaigns, and days before Anthropic’s chief executive publicly called on the industry to slow down. Taken together, the two disclosures mark the moment AI-enabled offensive cyber capability stopped being a hypothetical red-team scenario and became a parameter labs must manage in production. Governments and infrastructure operators, not just the labs themselves, will now have to decide what “managing” actually means.