Capsule Security fine-tunes Nvidia Nemotron models to stop rogue AI agents

Capsule Security, an agentic artificial intelligence security startup, has unveiled a detection system utilizing Nvidia Corp.'s Nemotron models, fine-tuned specifically for the purpose. This system, dubbed the "AI circuit breaker," operates by judging an agent's intended action in real-time before execution. Customers can subsequently allow, flag, or block the action in real-time, thereby establishing a control layer external to the agent.

This mechanism is designed to address the increasing prevalence of agents with access to sensitive data, source code, or production infrastructure. The system utilizes two of Nvidia's largest Nemotron models, Nemotron 3 Ultra, and Nemotron 3, which are capable of running within an agent's workflow without significantly hindering performance.

According to Capsule Security, the system achieved 98% accuracy on the StepShield benchmark, a metric for detecting rogue agent behavior at the step level. The benchmark utilized 9,429 code-agent trajectories derived from real incidents. While the startup acknowledges that rule-based guardrails can be effective, they often result in a high number of false positives.

In contrast, the fine-tuned models used by Capsule can classify a narrow set of actions with high precision and speed, making real-time decisions in as little as 71 milliseconds. The company claims that billions of tokens across millions of agent interactions are currently processed using this technology, with customers ranging from financial institutions to technology companies.

The startup's co-founders, Naor Paz and Lidan Hazout, emphasize that the core AI security risk now lies not just in what people can do with agents, but in what autonomous agents decide to do on their own. The release of this system follows the disclosure of two prompt injection vulnerabilities in widely used AI tools by Microsoft and Salesforce, both of which have since been patched.