⚠ INJECTION DETECTED
click to scroll down

// AI SECURITY

Stop prompt injection.

In 23 milliseconds.

Self-hosted protection that outperforms GPU models 8x its size. No external dependencies.

terminal

$ curl -X POST /scan

Response:

{

“malicious”: true,

“score”: 0.97,

“latency_ms”: 23

}


// PROTECTING USERS OF

OpenAI
Claude
Open WebUI
OpenClaw
GitHub Copilot
Custom Apps
0.998
F1_SCORE
23ms
LATENCY
355MB
RAM
CPU
NO_GPU

AI agents are under attack.

Hidden Instructions

White text in documents. Invisible to humans, visible to AI.

Zero-Click Exploits

Microsoft Copilot exfiltrated data with no user interaction.

Config Won’t Save You

System prompts and allowlists are insufficient.


Multi-layer detection pipeline.

INJECTION DETECTED
1. Statistics
2. Heuristics
3. Semantic
4. Neural
Safe
Layer 4 fallback: Meta Prompt Guard 2 · Built with Llama

Built for production.

23ms latency

Imperceptible in production

Self-hosted

No data leaves your infra

CPU-only

No GPU required

48+ languages

Global coverage


Prompt Guard vs. the alternatives.

Not production-ready

NVIDIA’s own docs: “not recommended for production without further customization and testing.” Most teams deploy it anyway.

High bypass rate: emoji and characters

Instructions hidden in emoji or invisible Unicode bypass NeMo every single time. Independent research also measures a 72.54% bypass rate on character injection. (Source: arxiv 2504.11168)

500ms latency and GPU required

1 to 3 extra LLM calls per message, GPU recommended. Prompt Guard: 23ms on standard CPU, no extra hardware.

Open source on GitHub

The full source code is publicly available. You can inspect how it works, fork it, and adapt it to your own setup.

Sources: NVIDIA NeMo Docs: docs.nvidia.com/nemo/guardrails · arxiv.org/abs/2504.11168 (LLMSec Workshop 2025)


Works with your stack.

Claude Code
hooks
OpenAI API
proxy
GitHub Copilot
extension
Open WebUI
filter
LLM Gateway
middleware
Any API
REST

Ready to secure your AI?

Get in touch for a demo or technical discussion.


Frequently asked questions.

No. The model is trained on public data. We have verified that this data does not correspond with known benchmarks. Any overlapping data has been fully removed from the framework, so there is no overlap between training data and benchmarks.

Among other methods, through AI red teaming: our own AI actively tries to break through the framework. The results of these attempts are used to improve the model. This is a continuous process where the framework sharpens itself over time.

No. We see Prompt Guard as an essential layer that every AI application should use, but prompt injection can never be fully mitigated. It remains possible that attacks get through the framework. Prompt Guard significantly reduces the attack surface, but does not replace a broader AI security strategy.

Yes. We can custom train models and frameworks for specific use cases, specific system prompts, and industry-specific threat patterns. We also offer general AI security advice. Get in touch to discuss the options.

On average a few tens of milliseconds per request. Prompt Guard works in layers: as soon as one of the detection layers gets a hit, you receive a result immediately without processing all remaining layers.

Prompt Guard is on-premise only. You run it entirely within your own infrastructure, without any dependency on external services.

No. Prompt Guard runs locally within your infrastructure, prompts never leave your environment.

Because Prompt Guard runs fully on-premise, you are in control of all data. If your environment is GDPR-compliant, so is Prompt Guard. LTech Consultancy keeps all business data within the Netherlands for all engagements.

Yes. The model is trained on attack prompts in multiple languages. Multilingual coverage is continuously expanded.

Yes. Prompt Guard works as model-agnostic middleware and inspects the prompt before it reaches your LLM. It works with OpenAI, Google Gemini, Anthropic Claude and any other model.

The model is updated daily to weekly. Updates are rolled out periodically without any changes to the API or integration on your end. Known attack techniques or vulnerabilities can be submitted directly, which can lead to an immediate model improvement.