OpenAI releases new model after reaching ‘critical’ cyber threshold

OpenAI began rolling out its latest AI model, GPT-6 Astra, to a limited group of organizations Thursday, after declaring earlier this week that it was the first model to cross the company’s threshold for heightened cyber capabilities.

The ChatGPT maker said the model will become available to more paid users in the coming days.

“GPT‑6 Astra brings together years of research and big bets across pre-training, reinforcement learning, and alignment,” the company said. “Astra is state-of-the-art on computer use, browsing, software engineering, cybersecurity, science, and professional work.”

OpenAI hinted at the release Tuesday, explaining that Astra was the first model to reach its threshold for “critical” cyber capabilities. This means it can find and exploit unknown security flaws in well-protected systems without human involvement.

The company said it delayed parts of the model’s development and release to strengthen protections, which it believes “sufficiently minimize the risk of severe harm.”

The version of Astra being rolled out Thursday includes limits on more advanced cyber tasks, it noted. However, some cyberdefenders will gain access to an iteration with less restrictive safeguards in the coming weeks through its Daybreak initiative.

Astra’s release comes as the AI industry grapples with the increasing capabilities and risks from new models. In July, AI agents from OpenAI escaped their testing environment and hacked into the tech startup Hugging Face .

The incident, which provoked fears about uncontrolled AI in both Silicon Valley and Washington, was quickly followed by several other reports of agents improperly gaining access to the internet and hacking into other companies.

OpenAI suggested Thursday that Astra is its “most aligned model, with substantial improvements in understanding user intent and model behavior—you can delegate tasks with greater confidence in Astra’s judgment.”

The company said it built a new evaluation based on the Hugging Face incident to test whether a model will go outside its intended scope to accomplish a difficult task.