Capsule Security Launches AI Circuit Breaker to Stop Rogue Agents

Capsule Security, founded in 2025 by CEO Naor Paz and CTO Lidan Hazout, has unveiled what it calls the first dedicated runtime security layer for autonomous AI agents. Announced on September 2, 2026, the product—branded an "AI circuit breaker"—is engineered to detect, evaluate, and block rogue agent behavior in real time, before any irreversible action takes place. The founders launched the firm after identifying a critical security gap created by agents that can reason, call tools, and execute actions on their own, including operating infrastructure and interacting with other systems that often house credentials (check yours with a password checker) and sensitive network endpoints (scan them using a port scanner).

The core challenge Capsule addresses is speed. Traditional pre-execution review of agent behavior is rarely viable at production scale because the latency introduced is often greater than the action itself, making such oversight both slow and expensive. Capsule's platform uses its own specialized detection AI trained on NVIDIA Nemotron 3 Ultra, leveraging real agent traces, human review, and adversarial examples to teach its models the precise boundary between authorized and rogue behavior. The approach deliberately avoids routing traffic to large general-purpose models for every decision, keeping both cost and latency manageable.

Benchmark results are striking. Capsule's more accurate model achieved 96.9% detection accuracy compared with 86% for the strongest third-party model evaluated, while its decision time can drop to as little as 71 milliseconds—fast enough to sit inline within an agent's workflow without adding friction. The company also cut the larger model's memory footprint by nearly 50%, allowing the evaluator to run directly inside the agent execution path and intervene milliseconds before an action commits. "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," Paz said at launch. "Human trust in AI depends on our ability to stop that action before it happens."

Capsule claims 98% efficiency for the circuit breaker's decision maker in testing, positioning the product as a proactive control layer rather than a forensic one. The evaluator can allow, flag, or block an intended action in real time, providing an independent checkpoint against agents that access sensitive data, write code, modify infrastructure, or communicate with external services. With post-incident monitoring only surfacing damage after the fact, Capsule is betting that runtime interception will become the standard model for governing autonomous systems—particularly as enterprises expose more production environments (audit yours with an SSL/TLS checker) to agent-driven workflows.