Capsule Security has released a detection system that fine-tunes two Nvidia Nemotron models to assess an AI agent’s behaviour just before execution and allow, warn or block it in real time, SiliconANGLE reported on Wednesday.

The system focuses on implementing an external control layer that screens in advance whether an agent’s assigned task matches the action it is about to carry out. It targets a situation in which more agents have access rights to sensitive data, source code and operational infrastructure.

The company said existing access controls and approval procedures only limit the scope of what agents can access. They cannot determine whether each individual action aligns with the instructed task. Post-incident monitoring can also identify problems only after damage occurs.

Capsule Security said its system achieved 98 percent accuracy on the StepShield benchmark and detected violations at the stage they occurred. StepShield evaluates step-by-step deviation detection performance based on 9,429 code-agent execution paths drawn from real incidents.

Capsule Security used Nemotron 3 Ultra, the largest model in Nvidia’s open-source Nemotron 3 family, to train its models. The training data included real agent execution logs and adversarial examples marking boundaries of permitted actions, which were reviewed by humans.

Of the two models Capsule Security introduced, the larger model can run on a single Nvidia L40S GPU by cutting memory requirements by nearly half without performance degradation, the company said.

Naor Paz (나오르 파즈), Capsule Security co-founder and CEO, said the core AI security risk goes beyond what humans can do with agents and lies in what autonomous agents decide to do on their own. "Once software can reason and take actions using tools, a wrong decision can turn into a real incident in seconds," he said.