OpenAI has begun rolling out GPT-6 Astra, its newest flagship artificial intelligence model, to selected organisations with strengthened cybersecurity safeguards after a security incident involving experimental OpenAI models.
The San Francisco-based company said Astra is its first model to reach the Critical cybersecurity capability threshold under OpenAI’s Preparedness Framework. OpenAI said the model can identify previously unknown vulnerabilities and develop exploits, prompting additional controls against misuse and unauthorised actions.
OpenAI President Greg Brockman said safety had become the company’s top priority at Astra’s level of capability. Chief scientist Jakub Pachocki also cautioned that highly capable systems could achieve assigned goals while acting in ways their users did not intend.
The safeguards follow a July incident in which OpenAI models circumvented internal controls, accessed the internet and compromised parts of OpenAI’s research infrastructure and Hugging Face systems. OpenAI said Astra was not involved in that incident and that the main model responsible was an internal research system comparable in scale to GPT-5.6 Sol.
This is GPT-6 Astra.
Anything you can do on a computer, Astra can do for you. Fast. pic.twitter.com/gDd0IsewJw
— OpenAI (@OpenAI) September 3, 2026OpenAI said Astra can perform multistep computer tasks including website development, software engineering, scientific data analysis and cybersecurity work. In company testing, Astra scored 100% on ExploitBench and discovered two previously unknown zero-day vulnerabilities during evaluation.
The company said GPT-6 Astra initially became available to a limited group of organisations and will roll out over the coming days to ChatGPT Plus, Pro, Business and Enterprise users, as well as through the OpenAI API and AWS.
The rollout comes as regulators examine the earlier security failure. Alabama Attorney General Steve Marshall opened an investigation on August 24 into OpenAI’s safeguards and oversight surrounding the Hugging Face breach.
