GPT-6 Astra released; Critical cyber classification alongside higher computer-use capability

OpenAI has begun a staged release of Astra, disclosed on September 3 with a 72-hour extension.

Vendor-reported results put OSWorld at 72.6 per cent, Terminal-Bench 4 at 57.9 per cent and ExploitBench at 100 per cent; these are not independent reviews.

The safety overview flags a Critical cybersecurity threshold, and chain-of-thought monitorability fell in some adversarial tests.

A least-privilege sandbox, approvals and immutable logs are essential.