Llama-MCP-Prompt-Guard-2-V3
a classifier protecting MCP tools against prompt injection
2 min readupdated
This model is a fine-tuned version of Meta's Llama Prompt Guard 2 classifier (86M parameters), specifically optimized to analyze English tool descriptions within the Model Context Protocol (MCP) framework. Operating as a binary classifier, it flags prompt injection and jailbreak attempts hidden in tool descriptions. Designed as an asynchronous advisory signal (SHADOW advisory signal), it supplements rather than replaces core security controls. It is built for developers working with MCP architectures and agentic tool-calling systems.
Key features
- Base model: a fine-tuned variant of Meta Llama Prompt Guard 2 86M.
- Purpose: text classification to detect prompt injections and jailbreaks.
- Use case: scanning tool descriptions across the MCP ecosystem.
- License: distributed under the Llama 4 Community License.
What you can do
- Integrate the model as a secondary security filter to screen incoming tool descriptions in MCP setups.
- Run the provided runtime files (model.safetensors, config.json, etc.) using commit pinning and SHA-256 verification for secure deployment.
- Set up asynchronous processing of advisory signals to audit the security of agentic environments.
Example
Load the model using standard Hugging Face libraries for binary text classification while verifying the SHA-256 hash of the weight files.
Read more
DetailsArchitecture: 86M-parameter binary classifier
Language: English (English MCP tool description)
License: Llama 4 Community License
Notes: the training dataset and raw data are not included in the repository










Comments