Astra will be OpenAI's first model with 'critical' cyber abilities
What's the story
OpenAI has announced its upcoming AI model, Astra, as the first to meet the company's "critical" cyber capabilities threshold. The tech giant plans to launch a public version of Astra soon. However, only select partners in its Daybreak Blue early-access program will get access to the model's advanced cyber capabilities at launch.
Cybersecurity concerns
Astra's critical cyber threshold
OpenAI defines a model as having reached its critical cyber threshold when it can autonomously discover and exploit unknown vulnerabilities in real-world software.
After discovering this capability in Astra, the company decided to halt further development until appropriate safeguards could be implemented.
This decision came after an incident where agents running two of its models exploited vulnerabilities in a supposed siloed testing environment.
Enhanced security
Model more resistant to jailbreaking attempts
Astra can find new software vulnerabilities and create ways to exploit them for hacking. It can even "chain" multiple exploits together, a technique used to dig deeper into a target system.
OpenAI has also made Astra more resistant to jailbreaking attempts. In tests, it successfully refused unsafe queries at a much higher rate than previous models.
Model comparison
Astra outperforms other AI models on cybersecurity benchmarks
OpenAI claims Astra outperforms industry-leading AI models like GPT-5.6 Sol and Anthropic's Mythos on cybersecurity benchmarks.
However, the company also warns that its misalignment monitor may sometimes flag legitimate activity as potential cyber misuse or unauthorized behavior, slowing down or stopping the model inadvertently.
Partners in OpenAI's Daybreak program will get early access to a less restricted version of Astra with more robust cyber capabilities.