According to OpenAI, the upcoming AI model Astra is making such significant progress in the field of cybersecurity capabilities that security measures are being tightened.
OpenAI writes in a blog post that its new model Astra, which is still under development, has reached a “critical cybersecurity threshold.” Internal evaluations indicate a possible threshold breach, leading the company to temporarily slow down development. Earlier models, such as GPT-5.6 Sol, were still assessed at a high, but not critical, threshold. According to OpenAI, the current situation surrounding Astra requires a stricter approach.
“Critical threshold”
According to the Preparedness Framework, which OpenAI established in 2023, critical cybersecurity is involved when a model can independently identify and develop zero-day exploits in various realistic systems without human intervention. The independent execution of new cyberattack strategies against hardened targets also counts.
Preliminary evaluations of Astra do not rule out reaching this critical threshold, according to OpenAI. The model was not linked to recent incidents, such as the exploitation of Hugging Face. OpenAI does report that performance is “strong enough to warrant increased vigilance.”
Security measures
OpenAI has subsequently ramped up internal security. This includes stricter access, isolated testing environments, advanced encryption, and real-time monitoring of all high-risk activities. Internal activities that do not meet these requirements are reportedly paused for the time being.
Additionally, universal monitoring is applied to all agentic applications of Astra, where high-risk actions are immediately detected and interrupted if necessary. Collaborations with governments and AI safety organizations are also being established to enable external testing under controlled conditions.
The rapid evolution of AI models creates new challenges in the field of cybersecurity. Previously, Claude managed to escape internal security tests, as did other models from OpenAI and Anthropic. OpenAI emphasizes that advanced cyber-capable AI models should primarily help defenders identify vulnerabilities faster than attackers.