OpenAI's new model, Astra, is what the company calls its most capable coding and reasoning system yet, and it ships with something new attached: the ability to find and exploit software vulnerabilities on its own, well enough that OpenAI briefed the Pentagon and CISA before release. That is a plain description of what happened this week, confirmed in OpenAI's own release notes and reported by Wired and TechCrunch on September 1. The capability is not a benchmark score in a paper. It is a model that can chain together the steps a human penetration tester takes, scanning a system, finding the weak point, writing the exploit, without a human at each step.

The gap is what Astra does not yet have: a deployment base of security teams actually running it as a defensive tool at the scale it could be run offensively. Anthropic's Claude and Google's Gemini already sit inside thousands of enterprise security operations centers, wired into SIEM platforms and patch-management pipelines, the products of contracts signed and compliance-cleared over the past two years. Astra has the exploit-generation skill first and the defensive install base second, and that ordering is the whole risk. A vulnerability scanner that only attackers have integrated is not a balanced tool, it is a head start, and OpenAI's own safety researchers said as much to The Verge on the same day the model shipped: the monitoring systems meant to catch misuse of Astra were not the ones stress-tested against a model this good at breaking in. The clock that matters now is not a training run. It is how fast CISOs at named enterprises, not hypothetical ones, get equivalent tooling under contract before the exploit side of that ledger gets used against them.