Capsule Security Adds a Last-Second Block for AI-Agent Actions
The new control layer judges a planned action before it runs, rather than relying only on access permissions or post-incident review. Its reported benchmark results are promising, but the company’s internal model comparison remains difficult to assess fully.
Listen to this story
The audio brief
Story brief
3 key pointsCapsule Security is now offering a runtime guardrail that judges an agent’s next step just before it executes, with customer-configurable allow, flag, or block outcomes. The system targets actions involving code, sensitive data, and infrastructure—areas where static permissions may miss context and monitoring arrives too late.
- 01
Capsule says its detectors classify planned actions rather than generate responses, helping keep latency low enough for execution-path controls.
- 02
Training used agent traces, human review, and adversarial examples with two fine-tuned Nvidia Nemotron models.
- 03
Memory use for the larger model was nearly halved without reported performance loss, enabling single-GPU deployment.
Capsule Security has released a runtime control layer that checks an AI agent’s intended action immediately before execution. Customers can allow, flag or block the step, creating a decision point for agents that can access sensitive data, write code or operate infrastructure.
A check before the action lands
The capability is available now. Capsule places the control outside the agent, where it evaluates a specific planned step rather than simply setting broad access rights. The company says permissions and approval workflows can restrict what an agent may touch, but may not determine whether a particular action fits its assigned task. Post-incident monitoring, by contrast, finds a problem after damage has occurred.
That is the basis for Capsule’s circuit-breaker framing. Its detectors perform a narrow classification task instead of generating a full response, a design the company says lets them sit in an agent’s execution path with little delay.
Two models trained for one judgment
Capsule built the detection system on two Nvidia Nemotron models fine-tuned for real-time action evaluation. Nvidia Nemotron 3 Ultra supported the training run, which combined agent traces, human review and adversarial examples.
- Agent traces
- Human review
- Adversarial examples
Capsule says decisions arrived in as little as 71 milliseconds. It also says it cut the larger model’s memory requirement by nearly half without a performance hit, allowing deployment on a single Nvidia L40S GPU.
The comparison still needs context
Capsule said the system reached 98% accuracy on StepShield, an academic benchmark containing 9,429 code-agent trajectories drawn from real incidents. Its strongest detector scored 96.9% in Capsule’s internal testing, compared with 86% for the strongest third-party general-purpose model it evaluated.
The two results answer different questions. Capsule did not identify the third-party model or provide a score breakdown for its internal comparison, limiting how fully that result can be assessed. Capsule also says billions of tokens across millions of agent interactions pass through its technology, with customers in financial services and technology.
Sources
- markets.businessinsider.comCapsule Security Builds “AI Circuit Breaker” for Rogue Agents with NVIDIA Nemotron
- siliconangle.comCapsule Security fine-tunes Nvidia Nemotron models to stop rogue AI agents - SiliconANGLE
Reader comments
Newest comments first. Replies stay oldest first.