OpenAI releases GPT-6 Astra with stronger computer use and a higher cyber-risk classification

OpenAI released GPT-6 Astra across a staged product and API roll-out. Its safety overview classifies the model at the company’s Critical cybersecurity threshold. OpenAI reported 72.6% on OSWorld, 57.9% on Terminal-Bench 4 and 100% on ExploitBench; these are vendor-reported results, not independent reproduction.

Why it matters: capability and risk move together. OpenAI also reported weaker chain-of-thought monitorability in some adversarial evaluations, reinforcing the need for external controls.

Impact: pilot behind least-privilege sandboxes, approval gates and immutable logs. OpenAI release · Safety overview · a report

report.com/legal/litigation/openai-launches-new-astra-model-amid-growing-scrutiny-over-agents-safety-2026-09-03/