OpenAI announced on September 1 that its upcoming Astra model reached "Critical" under the company's Preparedness Framework — the highest cybersecurity risk tier OpenAI has ever assigned to a model. Internal testing found Astra could discover novel software vulnerabilities and build working exploit chains largely without step-by-step human guidance.
In evaluations, Astra scored 100% on ExploitBench, a benchmark measuring exploit-development skill, and independently found two previously unknown vulnerabilities in real-world software. Testers said it built a full browser-compromise chain that escaped a sandbox and escalated privileges from a standard user to root/admin on a hardened operating system.
Before resuming development, OpenAI added chain-of-thought monitoring to halt unauthorized actions, tightened model-weight encryption, restricted network and tool access, and pushed refusal training that lifted the model's rejection rate on cyber-jailbreak prompts to 91.5%, up from 59% in its predecessor. A general version of Astra is expected to launch publicly soon, but the most powerful cybersecurity tools will stay limited to vetted testers in OpenAI's "Daybreak Blue" early-access program.