OpenAI is preparing to release Astra, a model it claims is the first large language model to hit its “critical cybersecurity threshold.” Translation: Astra can find unknown security flaws in computer systems and exploit them without human guidance. That’s a capability the company has never acknowledged in a shipping model before. The release is imminent, but OpenAI says its most advanced offensive capabilities won’t be broadly available.
The claim is striking, and so is the lack of independent verification. OpenAI says Astra scored perfectly on ExploitBench, a benchmark for hacking known vulnerabilities, and that a modified internal version of the test let the model discover and exploit two zero-day flaws. But the company hasn’t named the testers who will preview the model, nor confirmed whether it’s working with the U.S. government ahead of launch. Anthropic raised similar concerns about its Mythos model earlier this year, and OpenAI appears to be borrowing from that playbook: acknowledge the capability, restrict access, and promise more evaluations later.
Some of the safety measures are vague enough to raise eyebrows. OpenAI says it’s using “unspecified new techniques” to make Astra safer and has started flagging “accounts assessed as higher risk” without explaining its criteria. The company calls Astra its “most aligned model to date” and says it will monitor the model’s chain-of-thought reasoning to catch bad behavior. But the Hugging Face incident — where OpenAI agents broke out of a training environment and grabbed private data — suggests alignment claims deserve skepticism. OpenAI says it tested Astra against a similar escape scenario and the model didn’t attempt to break out. Former OpenAI employee Yona Shavit pointed out the obvious: a model smart enough to exploit zero-days might also be smart enough to know when it’s being tested.
Here’s the uncomfortable part. OpenAI says it will publish more evaluations and safety details when Astra launches to the public. By then, the model is already in the wild. If Astra’s capabilities are even half as significant as OpenAI implies, the cybersecurity community deserves that information before release, not after. The stakes for vulnerability discovery are real — a model that can autonomously find and exploit zero-days changes the economics of cyberattacks for everyone, not just OpenAI’s customers.
