OpenAI limits Astra’s strongest cyber tools at launch

OpenAI is set to launch a public version of its upcoming Astra AI model, but will temporarily limit its most advanced cybersecurity capabilities to vetted testers and partners. Astra has been classified as having "critical" cybersecurity functions under OpenAI's Preparedness Framework, meaning it can independently identify and exploit previously unknown vulnerabilities in real-world systems.

This classification applies only to the initial version released to a limited group of alpha testers. Broader access will be granted through Daybreak Blue, a program for authorized defensive cybersecurity work. The restrictions mean the consumer-facing version will not have Astra's full offensive-security potential. OpenAI may refuse, slow, pause, or stop certain requests if monitoring systems detect potential misuse or unauthorized activity.

In tests, Astra significantly outperformed GPT-5.6 Sol in vulnerability identification and exploit development, achieving a perfect score on the ExploitBench benchmark. During internal testing, Astra discovered two unknown vulnerabilities and used them as part of an exploit chain while bypassing browser sandboxes and elevating privileges to root access.

OpenAI delayed parts of Astra's development and release to enhance protections against malicious use and unauthorized model actions. Astra has been trained to refuse harmful cyber requests more reliably and is protected by stronger system-level controls, monitoring, and restrictions. OpenAI's cyber jailbreak evaluations show Astra refused 91.5% of disallowed requests compared to 59% for GPT-5.6 Sol.