Capsule Security Builds “AI Circuit Breaker” for Rogue Agents with NVIDIA Nemotron

NVIDIA Nemotron model, fine-tuned by Capsule, catches 98% of rogue agent behavior on independent benchmark, outperforms OpenAI, Anthropic, Google frontier models

Capsule Security announced a solution based on Nemotron small language models that acts as an “AI circuit breaker” to stop rogue AI agents at runtime, before they can wreak havoc on a company’s digital infrastructure.

Researchers achieved 98% accuracy on StepShield, an independent academic benchmark for measuring whether security systems can identify and stop rogue agent behavior before damage occurs. The system detected violations at the exact step they occurred, an important consideration for enterprises processing millions of agent actions. The benchmark was led by a former Stanford researcher, with researchers from Cornell and other leading institutions.

The models evaluate an agent’s intended action immediately prior to execution, giving organizations the ability to allow, flag or block it in real time. This creates an independent control layer for agents that can access sensitive data, write code, operate infrastructure and interact with other systems. Permissions and approval workflows can limit what an agent is supposed to do, but they cannot always determine whether an action is appropriate within the context of a specific task. Post-incident monitoring only identifies the problem after the damage has occurred.

“The defining AI security risk is no longer only what people can do with agents. It is what autonomous agents can decide to do by themselves,” said Naor Paz, CEO and co-founder of Capsule Security. “When software can reason, use tools and take action, a wrong decision can become a real-world incident in seconds. Human trust in AI depends on our ability to stop that action before it happens.”

Also Read: CIO Influence Interview with John Elliott, Cybersecurity Author Fellow at Pluralsight

Specialized AI for real-time intervention
Because the models perform a narrowly defined classification task instead of generating a full response, they can operate directly in the agent’s execution path with minimal delay. This gives organizations an AI “circuit breaker” that can intervene while an action is still preventable.

Capsule used NVIDIA Nemotron 3 Ultra to support the training process, which combined real agent traces, human review and adversarial examples designed to teach the models the boundary between authorized and rogue behavior.

Capsule fine-tuned two NVIDIA Nemotron models. Testing found that they could provide strong detection without the cost and latency of sending every agent action to a large general-purpose model:

• Outperformed every general-purpose model tested in its internal benchmark – Capsule’s most accurate detector scored 96.9%, compared with 86% for the strongest third-party model evaluated.
• Made decisions in as little as 71 milliseconds – Both models were fast enough to operate within an agent’s workflow without creating a significant delay.
• Reduced the infrastructure required for deployment – Capsule cut the larger model’s memory requirements nearly in half without affecting its performance, allowing it to run on a single NVIDIA L40S GPU.

Runtime controls for the enterprise
Capsule’s technology is already protecting billions of tokens across millions of agent interactions. Its enterprise customers include leading financial institutions and technology companies across various sectors.

“AI agents represent a fundamentally new security challenge: they can reason, use tools, and take consequential actions at machine speed. Capsule helps organizations monitor agent behavior in real time and stop unauthorized actions before they execute. This gives security teams the confidence to expand their use of agentic AI while maintaining the security, governance, and accountability their clients expect,” said Phillip Miller, Vice President & Global Chief Security Information Officer, H&R Block.