OpenAI Says Astra Has ‘Critical’ Cyber Capabilities, Plans Limited Release

OpenAI says its next artificial intelligence model, Astra, has reached a “critical” level in the company’s cybersecurity capability assessment, a milestone that prompted the company to limit access to its most advanced functions.

The company said Astra became the first model it has developed to reach that level. OpenAI had delayed the release after determining that it could not rule out the possibility of critical cybersecurity capabilities. It used the intervening weeks to strengthen safeguards and monitor the model’s reasoning and actions, with systems designed to stop potentially unauthorized activity automatically.

OpenAI defines the critical level as the ability either to find undisclosed vulnerabilities in hardened systems and develop ways to exploit them without human intervention, or to independently devise and execute a new cyberattack from start to finish against a target when given only a high-level goal. The company said Astra met the first condition.

According to OpenAI, expert testing showed that Astra found several vulnerabilities in a hardened operating system and combined them to move from ordinary user access to the highest administrator privileges. It also achieved a perfect score on a benchmark measuring the ability to develop attack code from vulnerabilities.

OpenAI said the safeguards reduce the risk of serious harm and that it plans to release Astra soon. Its advanced cybersecurity capabilities will initially be available experimentally to only a small number of users.

“

After collecting more evidence and conducting additional evaluations to assess the model’s capabilities, we determined that Astra meets the critical level of cybersecurity capability we have defined.