Capsule launches AI circuit breaker to block rogue agents

Capsule Security released an AI “circuit breaker” on Sept. 2, 2026. The tool evaluates an autonomous agent’s intended action immediately before execution and can allow, flag or block that action in real time.

Capsule was founded in 2025 by Naor Paz, chief executive, and Lidan Hazout, chief technology officer. The founders developed the system to address risks from autonomous agents that can reason, use tools and act without human intervention. Paz described the defining risk as what agents may decide to do independently and emphasized the need to stop a harmful action before it executes.

The product uses specialized small language models (SLMs) that run inside an agent’s execution path rather than sending every action to a large general-purpose model for review. Capsule trained the models using NVIDIA Nemotron 3 Ultra on datasets that included real agent traces, human-reviewed examples and adversarial cases designed to mark the boundary between authorized and unauthorized behavior.

Capsule built two models to balance speed and accuracy. In internal tests the more accurate model reached 96.9% detection accuracy, compared with 86% for the strongest third-party model the company evaluated. The in-line evaluator can reach a decision in as little as 71 milliseconds. Engineers reduced the larger model’s memory footprint by nearly 50% so it can run within an agent’s workflow without adding meaningful delay.

The company reported 98% efficiency for the circuit breaker’s decision maker on StepShield, an academic benchmark that measures whether systems can stop rogue agent behavior before damage occurs. Capsule describes the product as an independent control layer for agents that access sensitive data, write code, operate infrastructure or interact with other systems, enabling automatic blocking or flagging of risky actions prior to execution. The company contrasted this preventive approach with post-incident monitoring, which identifies problems after an incident has caused harm.

Organizations are using autonomous agents for tasks such as code changes, infrastructure management and data handling, increasing the need for real-time controls. Capsule released the product as a runtime safety layer that can be inserted into existing agent pipelines to intercept potentially harmful decisions.

Naor Paz wrote that “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.” He added, “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.”