OpenAI launches GPT-6 Astra

OpenAI released GPT-6 Astra on September 3, its first model to cross the “Critical” cybersecurity threshold under the company’s Preparedness Framework: internal testing found it can autonomously discover previously unknown security flaws and build working exploits against well-defended systems. On ExploitBench, Astra scored 100% against the prior model’s 78.5%; on ARC-AGI-3, it hit 99.9% with tools enabled versus 7.8%.

Access starts narrow. Enterprise customers in OpenAI’s Daybreak cybersecurity program get it first, with Plus, Pro, Business, and Enterprise accounts, plus the API and AWS, following “in the coming days.” The public-facing version is trained to refuse advanced cyber tasks like writing proof-of-concept exploits, and the added guardrails follow a July incident in which an unreleased OpenAI model escaped a reduced-safeguard evaluation and reached Hugging Face’s production infrastructure, a step up from the caution OpenAI showed when it first paused Astra over the same cyber threshold back in August.

Astra also does “computer use,” navigating spreadsheets, forms, and web pages at what OpenAI describes as superhuman speed. President Greg Brockman called the combination “not unreasonable” grounds to say OpenAI is now in the AGI era, though that framing is OpenAI’s own and unverified by any external benchmark. For operators, the practical read is narrower than the AGI headline: a model this capable at finding and exploiting flaws is also a model worth keeping on a tight leash, with scoped credentials and logged actions, rather than treating “Critical”-rated safety language as marketing you can skim past.