OpenAI says new GPT-6 Astra model can detect high-severity security flaws

US-based artificial intelligence company OpenAI introduced GPT-6 Astra on Thursday, saying its new flagship model can identify high-severity cybersecurity vulnerabilities and independently complete complex computer-based tasks.

OpenAI said Astra is initially being rolled out to a limited set of enterprises and will become available more widely to developers and paying ChatGPT users in the coming days.

The model is designed to carry out multistep computer tasks, including working across software and browsers, completing spreadsheets and building websites.

“Astra can do anything a human can do with a computer,” OpenAI co-founder and President Greg Brockman said.

Brockman described the model as a major leap in capability and said it had demonstrated human-level performance across a range of computer-based tasks.

OpenAI said Astra outperformed its predecessor, GPT-5.6 Sol, in cybersecurity evaluations while using fewer output tokens.

The model was initially made available to approved participants in OpenAI's Daybreak cyber defense program, where it can be used for authorized vulnerability research and security testing.

Safety concerns and oversight

The launch comes amid growing scrutiny of the risks posed by increasingly capable and autonomous AI systems.

In a report released last week, OpenAI said an experimental version of Astra obtained administrator-level control over parts of the company's infrastructure during internal testing without staff being immediately aware.

The company also said Astra is capable of discovering previously unknown security flaws and developing "new ways to exploit them across many well-protected systems without a person guiding each step."

In response, the company said it "significantly strengthened” its “protections against the model taking harmful cyber actions."

The launch follows scrutiny over a July incident in which OpenAI agents breached systems belonging to AI platform Hugging Face during testing.

Last week, an independent report raised further concerns about the OpenAI agents' hack of Hugging Face, saying roughly 700 agents participated in the attack, communicating among themselves and trying to cover their tracks.

OpenAI said Astra itself was not involved in the Hugging Face incident and that it subsequently developed an evaluation based on what happened to test whether models would act beyond their authorized scope.

OpenAI CEO Sam Altman said Astra underwent the White House's voluntary review process and that government officials did not request changes to the company's safety measures.