Astra crosses OpenAI’s Critical cyber threshold

Astra is the first OpenAI model to meet the Critical cybersecurity capability threshold of the company’s Preparedness Framework, OpenAI said in its September 1 post “Path to Astra” on the OpenAI site. The company says the release comes with stronger safeguards.

Testing Catalog reports that Astra scored 100% on ExploitBench. OpenAI then built a harder benchmark, “ExploitBench - Internal Port”, from 20 high-severity vulnerabilities in V8, the JavaScript engine inside Chrome, all disclosed more recently. On that set Astra reached “much higher arbitrary code-execution rates than GPT‑5.6 Sol”. Arbitrary code execution means the attacker gets to run their own code on the target, which is usually the goal of a browser exploit.

The evaluation also turned up new bugs. Astra found 2 new zero-day vulnerabilities and turned them into working exploit chains, Testing Catalog says. An exploit chain links several flaws into one attack that works end to end. CNBC reported that, according to OpenAI, Astra can find previously unknown security flaws and exploit them without step-by-step guidance from humans.

In August OpenAI stopped part of its internal work with Astra over the cyber threat the model could pose. It also described isolated test environments, extra monitoring and detection tooling for models it classifies as high-capability. “We have deployed universal monitoring of risky actions and misalignment across all agentic uses of Astra, including training and evaluation,” OpenAI said at the time.

OpenAI still plans to make Astra available “soon,” CNBC reported. Testing Catalog says the version that ships will have its cybersecurity capabilities limited.