Capsule Security Launches Runtime Control to Block Rogue AI Agent Actions
Summary
Capsule Security launched a runtime control that evaluates AI agents’ intended actions before execution. The company reports 96.9% detection accuracy for its leading model and 98% efficiency against the StepShield benchmark.
Key points
- Capsule Security says its evaluator runs in an agent’s execution path and can allow, flag, or block an intended action before it occurs.
- The company says its models were trained using real agent traces, human review, and adversarial examples, with NVIDIA Nemotron 3 Ultra supporting the training process.
- Capsule reports 96.9% detection accuracy for its more accurate model, compared with 86% for the strongest third-party model it evaluated.
- The company says decisions can be made in as little as 71 milliseconds and that it reduced the larger model’s memory requirement by almost 50%.
- Capsule claims 98% efficiency against StepShield, an academic benchmark for identifying and stopping rogue agent behavior.
Article Details
- Event Type
- AI security product launch
- Impact
- Capsule Security launched an AI circuit breaker designed to evaluate autonomous agents’ intended actions in real time and allow, flag, or block them before execution. Capsule reports 96.9% detection accuracy for its more accurate model, decisions in as little as 71 milliseconds, and 98% efficiency in testing against the StepShield benchmark. These are company-reported results; the article reports no real-world incidents or independently confirmed impact.
People
Vendors
Capsule SecuritySuch a ‘device’ has now been developed and released by Capsule Security.NvidiaCapsule’s solution uses its own specialized AI for real-time intervention. The firm used NVIDIA Nemotron 3 Ultra to support the training process, combining real agent traces, human review and adversarial examples