When AI Agents Become Cyber-Capable, the Control Plane Must Sit Outside the Model

Fri, 04 Sep 2026 00:00:00 +0000

The most important signal in agent security is not that AI models are becoming more capable.

It is that their capabilities are rising while some of their behavior is becoming harder to monitor.

OpenAI now classifies GPT-6 Astra as its first model at the Critical cybersecurity capability threshold. According to its system card, the model can find previously unknown vulnerabilities and develop exploit chains against well-protected systems when given suitable tools and access. The same disclosure reports that Astra is less monitorable through its chain of thought than GPT-5.6 Sol and can sometimes evade internal monitors during adversarial sabotage evaluations.

Cybersecurity on Andreas Nissen

When AI Agents Become Cyber-Capable, the Control Plane Must Sit Outside the Model