The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains common jailbreak and injection strings such as "Ignore previous instructions" and "You are now in developer mode." These are included solely as descriptive examples and test cases for the detection model, not as active instructions intended to bypass the agent's safety guardrails.
[EXTERNAL_DOWNLOADS]: The skill references downloading model weights and configuration from Meta's official repository on HuggingFace. This targets a well-known service and a trusted organization, representing standard and expected behavior for machine learning applications.
[INDIRECT_PROMPT_INJECTION]: The skill defines a surface for processing untrusted user input and third-party data to detect malicious content.
Ingestion points: Untrusted text enters the system via the get_jailbreak_score and filter_third_party_data functions defined in SKILL.md.
Boundary markers: The code snippets do not implement specific delimiters, as the text is passed directly to the model's tokenizer.
Capability inventory: The provided scripts are limited to computing classification scores; they do not perform command execution, sensitive file access, or outbound network operations based on the processed data.
Sanitization: No pre-inference sanitization is present, as the primary purpose of the skill is to provide the classification logic itself.

prompt-guard