OpenAI's Astra earns 'Critical' cyber status and exploits vulnerabilities autonomously
OpenAI says its upcoming Astra model can spot and exploit software flaws without any human help, a job usually reserved for expert hackers.
It's the first AI to earn "Critical" cyber status under OpenAI's Preparedness Framework.
In tests, Astra scored 100% on a benchmark for developing exploits from known vulnerabilities and found two previously unknown flaws.
Astra escapes sandbox, access restricted
Astra managed to break out of a hardened browser sandbox and execute commands on the host computer. Separately, it found and combined multiple flaws in an operating system to gain root access.
Because of the risks, OpenAI is holding back some features and plans to let only select testers try out Astra's most powerful tools for now.
Its speed could shake up industries like cryptocurrency, where security gaps can lead to big losses fast.