AI Circuit Breaker Catches 98% of Rogue Agents

AI circuit breaker technology from Capsule Security caught 98% of rogue AI agent behavior in an independent benchmark Wednesday, outperforming detection systems built on OpenAI, Anthropic and Google frontier models after fine-tuning on NVIDIA (NVDA) Nemotron.

Key Takeaways

Capsule Security’s AI circuit breaker caught 98% of rogue AI agent behavior in an independent benchmark
The system flags and halts autonomous agents before they complete harmful actions such as unauthorized data transfers or unapproved purchases
Capsule built its detection system on Nemotron, an open multimodal model family Nvidia designed for agentic reasoning
Capsule has not published the benchmark in full, leaving the test’s scope and dataset size unverified

Capsule detailed the results in a GlobeNewswire release. The system flags and halts autonomous agents before they complete harmful actions such as unauthorized data transfers or unapproved purchases.

Capsule did not disclose the exact detection scores posted by the OpenAI, Anthropic and Google models it says it beat.

How The AI Circuit Breaker Watches Agents In Real Time

AI agents are software systems allowed to act autonomously — sending emails, executing code or moving money rather than just answering questions. A rogue agent drifts from its instructions, through a manipulated prompt or an unintended decision, and carries out a step no human approved.

The tool takes its name from the electrical device that cuts power the instant a circuit senses a fault: it watches an agent’s actions and interrupts execution the moment it flags a rogue pattern.

Why Autonomous Agents Keep Slipping Off Script

Enterprise adoption of AI agents accelerated through 2026, with companies letting software handle onboarding, account management and developer integrations without a human checking every step. OpenAI found that firms including Basis, Clay and Exa Labs lean on agents for this kind of unsupervised task chain — the same work that creates room for an agent to go rogue.

The Benchmark Nvidia’s Nemotron Made Possible

Capsule built its detection system on Nemotron, an open multimodal model family Nvidia designed for agentic reasoning across graduate-level science, advanced math and visual understanding.

Fine-tuning that base let Capsule specialize the model for spotting anomalous agent behavior rather than general question answering.

What Enterprises Still Don’t Know About Agent Safety

Capsule’s 98% figure comes from one independent benchmark the company has not published in full, leaving the test’s scope and dataset size unverified. Whether the system holds up against novel jailbreaks — rather than behaviors it was tuned to catch — remains untested in public.