OpenAI rates Astra critical for cyber risk, ships it anyway

OpenAI classified its new model, Astra, as Critical for cybersecurity on September 1, per its Path to Astra report. Critical is the company's own Preparedness Framework threshold. It means Astra can find and build a working zero-day exploit in a hardened real-world system without a person in the loop.

OpenAI is restricting access to Astra's strongest capabilities rather than holding the model back. It says Astra refuses 91.5% of disallowed cyber requests in its own tests, against 59% for GPT-5.6 Sol.

OpenAI paused its largest frontier training run for two weeks after the Hugging Face breach. It restarted the run on August 28 under new safety requirements.

Set that against a September 1 report from The Information, relayed by Techmeme. Astra reportedly uses recurrent depth, looping the same block of transformer layers over a hidden state. That lets it spend extra compute on hard problems without writing the reasoning out as text.

OpenAI hasn't confirmed the architecture. Its chain-of-thought monitor, the one it says would have flagged the Hugging Face breach a day early, works by reading tokens. Reasoning that stays inside a hidden state gives it nothing to read.

Zvi Mowshowitz's postmortem calls that a symptom fix. He argues surveillance-based containment loses because offense beats defense, and wants mandatory third-party audits and public misalignment data instead of voluntary disclosure.

None of this changes what a solo developer runs on September 1. It matters for what gets allowed eighteen months out, once reading the reasoning stops being an option.