Silicon Angle reports that Capsule Security has released a new detection system designed to act as an "AI circuit breaker" for potentially rogue AI agents. This system utilizes two fine-tuned Nvidia Nemotron models to evaluate an agent's intended actions before execution, allowing customers to approve, flag, or block them in real time.
Capsule's system assesses whether a specific action aligns with the agent's assigned task, providing a control layer outside the agent itself. On the StepShield benchmark, the system achieved 98% accuracy in detecting violations at the step they occurred. The models are designed for a narrow classification job, enabling them to operate within an agent's workflow with minimal delay, with decisions returning in as little as 71 milliseconds.
Capsule's fine-tuned model reportedly outperformed frontier systems from OpenAI, Anthropic, and Google. The technology, trained on real agent traces and adversarial examples, can run on a single Nvidia L40S GPU. Capsule's customers include financial institutions and technology companies.
Source: Silicon Angle
