# Agent Guard Numbers: Speed, Accuracy and Cost

> Kriton's measured numbers with the method beside each, and a sourced comparison with other prompt and agent guards on speed, accuracy and cost.

Source: https://kriton.ch/numbers

Numbers

# Hold us to the numbers. We publish ours.

As of 16 Nov 2026

01 Side by side

## Our figures beside the ones vendors publish. Each with its source.

Attacks caught Higher is better

- Kriton Shield Kriton 100% 14 of 14 attack prompts, every one of 5 runs. Run file on request
- LLM Guard Protect AI 99.7% Recall on 20,000 prompts. [Source](https://huggingface.co/protectai/deberta-v3-base-prompt-injection-v2)
- NeMo Guardrails NVIDIA 98.9% With three safety models, 215 in-house cases. [Source](https://developer.nvidia.com/blog/measuring-the-effectiveness-and-performance-of-ai-guardrails-in-generative-ai-applications/)
- Lakera Guard Check Point 95.2% Its own PINT benchmark. [Source](https://github.com/lakeraai/pint-benchmark)

Not published: Prompt Shields, Bedrock Guardrails, Model Armor, Lasso for Claude Code, Prisma AIRS for Claude Code, Snyk Studio hooks, AI Guard for Claude Code. Each vendor tests on its own set, so read these as published claims side by side.

Time added to a prompt Lower is better

- Lakera Guard Check Point Under 40 ms Short prompts, setup not published. [Source](https://www.lakera.ai/blog/lakera-guard-fall-25-adaptive-at-scale)
- LLM Guard Protect AI 104 ms On a CPU; 7.65 ms on a GPU. [Source](https://protectai.github.io/llm-guard/input_scanners/prompt_injection/)
- Kriton Shield Kriton 341 ms Median of 210 decisions, our side only. Run file on request
- NeMo Guardrails NVIDIA 530 ms Average added with three safety models. [Source](https://developer.nvidia.com/blog/measuring-the-effectiveness-and-performance-of-ai-guardrails-in-generative-ai-applications/)

Not published: Prompt Shields, Bedrock Guardrails, Model Armor, Lasso for Claude Code, Prisma AIRS for Claude Code, Snyk Studio hooks, AI Guard for Claude Code. Ours is the decision on our side, before network time. Vendors measured on their own setups.

Ordinary prompts blocked Lower is better

- Lakera Guard Check Point Under 0.5% Calibrated production setups. [Source](https://docs.lakera.ai/docs/api/guard)
- Kriton Shield Kriton 6% 6 of 100; 5 are one prompt with the word password. Run file on request

Not published: Prompt Shields, Bedrock Guardrails, Model Armor, NeMo Guardrails with NemoGuard NIMs, LLM Guard, Lasso for Claude Code, Prisma AIRS for Claude Code, Snyk Studio hooks, AI Guard for Claude Code. We publish ours even where it is not the best number on the chart.

02 How we measured

## Our measured numbers

- Attack prompts stopped: 14 of 14 Every run, 5 runs of 14 attack prompts, as shipped, 8 Nov 2026. Run file on request.

- Prompt decision, server side: 341 ms Median of 210 decisions (42 prompts, 5 runs each); 582 ms p95; none over 3 s; 8 Nov 2026. Time on our side, before network and start-up on the person’s machine. Run file on request.

- Ordinary prompts blocked: 6 of 100 20 ordinary prompts, 5 runs each, as shipped; 5 of the 6 are one ordinary prompt with the word password in it; 8 Nov 2026. Run file on request.

- When Kriton cannot answer: Your setting decides You choose to let short prompts through or stop them. 0 errors in 210 decisions, 8 Nov 2026. Run file on request.

03 How we compare

## Before the action, or after it. That decides what a guard can stop.

| Product | When it acts | In Claude Code | Latency | Detection | False blocks |
| --- | --- | --- | --- | --- | --- |
| Kriton Shield Kriton Agentic Security | On the prompt before the agent sees it, and before each agent action runs | Yes | Published, above | Published, above | Published, above |
| Lakera Guard Check Point AI Guardrails (Lakera) | When your code calls the API: before the model call, after the response, or both, including tool calls Source, 16 Nov 2026 | No published integration; API call from your code Source, 16 Nov 2026 | Published Source, 20 Oct 2025 | Published Source, 2 May 2025 | Published Source, 16 Nov 2026 |
| NeMo Guardrails with NemoGuard NIMs NVIDIA | Input, dialog, execution and output rails, tools included Source, 16 Nov 2026 | No published integration Source, 16 Nov 2026 | Published Source, 3 Mar 2025 | Published Source, 3 Mar 2025 | Not published by the vendor Source, 16 Nov 2026 |
| LLM Guard, prompt injection scanner Protect AI (Palo Alto Networks) | On the prompt, wherever your code runs the scanner Source, 19 May 2025 | No published integration Source, 16 Nov 2026 | Published Source, 19 May 2025 | Published Source, 16 Nov 2026 | Not published by the vendor Source, 16 Nov 2026 |
| Lasso for Claude Code Lasso Security | Enterprise: on the request and before each tool call. Open-source hook: after the tool runs, warns without blocking Source, enterprise, 16 Nov 2026 Source, open source, 16 Nov 2026 | Yes, via Claude Code hooks Source, 16 Nov 2026 | Not published by the vendor Source, 16 Nov 2026 | Not published by the vendor Source, 16 Nov 2026 | Not published by the vendor Source, 16 Nov 2026 |
| Prisma AIRS for Claude Code Palo Alto Networks | Before the model processes a prompt or tool call, and on responses and tool outputs Source, 19 Nov 2025 | Yes, via Claude Code hooks Source, 19 Nov 2025 | Not published by the vendor Source, 16 Nov 2026 | Not published by the vendor Source, 16 Nov 2026 | Not published by the vendor Source, 16 Nov 2026 |
| AI Guard for Claude Code Zscaler | On the prompt, before MCP and web tools run, and on their responses Source, 16 Nov 2026 | Yes, via UserPromptSubmit, PreToolUse and PostToolUse hooks Source, 16 Nov 2026 | Not published by the vendor Source, 16 Nov 2026 | Not published by the vendor Source, 16 Nov 2026 | Not published by the vendor Source, 16 Nov 2026 |

04 Why before matters

## 247 days to find and contain a breach. Kriton stops the action before it runs.

247 days (183 to identify, 64 to contain), up from 241. [IBM, Cost of a Data Breach Report 2026](https://assets.ibm.com/adobe/assets/urn:aaid:aem:21111142-1251-4369-86fb-57b82f5bb108/original/as/Cost%20of%20a%20Data%20Breach%20Report%202026.pdf), published 10 Aug 2026.

## Bring your own prompts. We run the bench live.

[Book a demo](https://calendly.com/yc-cofounder/co-founder-meeting)[Request a pilot](https://kriton.ch/#pilot)
