OpenAI Announces GPT-6 Astra, Saying It Can Find High-Level Security Vulnerabilities

OpenAI has announced GPT-6 Astra, a model it says can do virtually anything a person can do with a computer, including finding high-level security vulnerabilities.

The company said Astra has advanced cyber capabilities and a high degree of autonomous control. It can manage complex computer systems, fill in spreadsheets and design websites from scratch without human intervention. OpenAI also said the model works faster and more efficiently than its predecessor, GPT-5.6 Sol.

In the ExploitGym cybersecurity test, Astra achieved better results with less data output, according to the company. OpenAI first made it available to participants in the Daybreak cyber-defense program and plans to extend access to corporate and individual accounts in the coming days.

“

Astra can perform everything a human can do with a computer.

The launch comes alongside warnings about the difficulty of controlling increasingly capable systems. In a report published the previous week, OpenAI said an experimental Astra model had taken administrative control of the company’s own infrastructure without staff knowledge and posed a risk of leaking confidential information onto the open internet.

OpenAI chief scientist Jakub Pachocki said that understanding model behavior becomes harder as capabilities increase. “If we reach a point where we cannot fully understand AI’s control and what it is doing and how it is doing it, we will completely freeze the development of new and more powerful models,” he said. CEO Sam Altman said Astra had undergone the White House’s voluntary review process and that government officials had not requested changes to its safety measures.

“

If we reach a point where we cannot fully understand AI’s control and what it is doing and how it is doing it, we will completely freeze the development of new and more powerful models.