OpenAI Begins Rolling Out Cyber-Critical GPT-6 Astra
OpenAI has begun a phased rollout of GPT-6 Astra, with tighter safeguards and additional restrictions on its most advanced cybersecurity capabilities.
A critical-capability model begins rolling out
OpenAI began rolling out GPT-6 Astra on September 3, two days after announcing that it had reached the “Critical” cybersecurity threshold under the company’s Preparedness Framework. It is the first OpenAI model to receive that classification. According to OpenAI, Astra can, with appropriate tools and access, discover previously unknown vulnerabilities and develop exploits across many well-protected systems without a person guiding every step.
The launch supersedes OpenAI’s earlier description of Astra as forthcoming, but availability is phased rather than universal at once. OpenAI says the rollout began with a limited set of organizations and is expanding over the following days to ChatGPT Plus, Pro, Business and Enterprise users, as well as the OpenAI API, Microsoft Azure and AWS Bedrock. Enterprise access is off by default unless an administrator enables it. The company calls Astra its most capable broadly deployed model and has published a safety overview and full system card.
The version entering general products refuses more advanced cybersecurity tasks such as producing proof-of-concept exploits. OpenAI says it plans to expand less restrictive access for approved defensive workflows—including vulnerability validation, malware analysis and detection engineering—through its Daybreak program in the coming weeks.
OpenAI says it delayed parts of Astra’s development and release while strengthening safeguards. Measures include stricter isolation, checkpoint encryption, monitoring of complete action trajectories, alignment checks before internal use, stronger refusals of harmful cyber requests and systems capable of stopping potentially unauthorized activity. The company also incorporated lessons from the OpenAI–Hugging Face infrastructure incident, although Astra itself was not used in that episode.
The newly published material adds evidence absent from the initial announcement. OpenAI reports that Astra scored 100% on ExploitBench and discovered two previously unknown V8 vulnerabilities while working through an internal benchmark. Those results remain developer-reported, and sensitive details are necessarily withheld while disclosure to maintainers proceeds. OpenAI also warns that added safeguards may slow, pause or stop legitimate work when activity is flagged as possible misuse or unauthorized behavior.
Why it matters
A frontier developer is now distributing a model it says can autonomously find and operationalize unknown weaknesses, while placing the most sensitive workflows behind additional controls. The phased launch shifts Astra from a prospective safety case to a live test of whether cyber-critical capability can be distributed without becoming a broadly reusable offensive service. Independent scrutiny will focus on both real-world capability and the rate at which safeguards interrupt legitimate defensive work.