Capsule Security Ltd. has released a real-time detection system built on two fine-tuned Nvidia Nemotron models to prevent rogue AI agent actions. Operating as an external control layer, the system evaluates agent actions in as little as 71 milliseconds, achieving 98% accuracy on the StepShield academic benchmark. Founded in 2025 by Naor Paz and Lidan Hazout, the startup previously raised $7 million in seed funding and aims to mitigate autonomous agent risks.
Capsule Security agent detection
- ▪Capsule Security Ltd. was founded in 2025 by Naor Paz and Lidan Hazout.
- ▪Capsule Security Ltd. released a detection system built on two fine-tuned Nvidia Corp. Nemotron models on September 3, 2026.
Fine-tuned Nvidia Nemotron models
- ▪Capsule Security Ltd. trained its detection models using Nvidia's Nemotron 3 Ultra model, incorporating real agent traces and human-reviewed adversarial examples.
- ▪Capsule Security Ltd. reduced the memory requirements of its larger fine-tuned model by nearly half, enabling it to run on a single Nvidia L40S graphics processing unit.
StepShield benchmark performance
- ▪Capsule Security Ltd.'s detector achieved 96.9% accuracy on an internal benchmark, outperforming frontier systems from OpenAI, Anthropic, and Google.
- ▪On the StepShield academic benchmark, Capsule Security Ltd.'s detection system achieved 98% accuracy in catching rogue agent behavior at the step where it occurred.
Real-time agent action control
- ▪Capsule Security Ltd.'s detection system evaluates an AI agent's intended action in real time, allowing customers to allow, flag, or block the action before execution.
- ▪Capsule Security Ltd.'s detection system processes decisions in as little as 71 milliseconds, allowing it to run within an agent's workflow without significant delay.
Capsule Security seed funding
- ▪Capsule Security Ltd.'s technology has processed billions of tokens across millions of agent interactions for customers including financial institutions and technology companies.
- ▪Capsule Security Ltd. launched publicly in April 2026 with $7 million in seed funding led by Lama Partners.
Prompt injection vulnerability disclosures
- ▪Capsule Security Ltd. disclosed prompt injection vulnerabilities in Microsoft Copilot Studio and Salesforce Agentforce on the day of the startup's public launch in April 2026, which have since been patched.
- ▪Capsule Security Ltd. co-founder Naor Paz stated that the defining AI security risk is what autonomous agents can decide to do by themselves.
Debatable claims
- ▪Autonomous agent decision-making is the primary security threat in enterprise AI
- ▪Enterprises should require external real-time monitoring for autonomous AI agents
- ▪Enterprises should deny autonomous AI agents access to sensitive production infrastructure
Story comments
Loading comments…