OpenAI now classifies the forthcoming Astra model at the Critical cybersecurity threshold—the first OpenAI model receiving that designation. OpenAI reports that Astra can identify unknown vulnerabilities and develop exploits across protected systems, while access to its strongest cyber capabilities will initially be restricted. OpenAI — 1 September 2026

• What changed: Astra moved from “possibly critical” to a confirmed Critical classification after additional evaluations.
• Why it matters: Agent security must now account for models capable of discovering and exploiting vulnerabilities with limited human guidance.
• Practical takeaway: Treat coding agents as privileged operators: isolate execution, restrict network/tool access, maintain command logs and require approval for destructive or security-sensitive actions.
• Analysis: OpenAI’s benchmark results and safety conclusions are vendor-reported; independent evaluation remains important.