The Agent Skills Directory

[SAFE]: The skill serves as documentation and a resource for implementing Meta's Prompt Guard security model. All code examples are standard implementations using the transformers and torch libraries.\n- [PROMPT_INJECTION]: The static analysis flags for prompt injection are triggered by example strings used to demonstrate the detection capabilities of the model (e.g., 'Ignore previous instructions', 'You are now in developer mode'). These are benign examples for educational and testing purposes and do not represent a threat to the agent executing the skill.\n- [EXTERNAL_DOWNLOADS]: The skill references models hosted on Hugging Face (meta-llama/Prompt-Guard-86M). As Hugging Face is a well-known and trusted repository for machine learning models, and the model belongs to a recognized organization (Meta), these downloads are considered safe and standard for the stated purpose of the skill.\n- [INDIRECT_PROMPT_INJECTION]: While the skill is designed to handle untrusted third-party data, it specifically provides sanitization and filtering logic to mitigate indirect injection risks rather than creating a vulnerability surface.

prompt-guard