The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted input text for analysis and interpolates it directly into a prompt for LLM evaluation, creating a vulnerability to malicious instructions embedded in the analyzed content.
Ingestion points: The text input variable processed by the local analysis functions and the LLM judge pipeline in SKILL.md.
Boundary markers: The skill instructions suggest informing the LLM that the text is data, but the provided code snippet uses simple string interpolation without robust delimiters like XML tags or clear separators.
Capability inventory: Network communication and model interaction through the @anthropic-ai/sdk library.
Sanitization: No input sanitization or filtering is performed on the analyzed text before it is sent to the LLM judge.
[INDIRECT_PROMPT_INJECTION]: The metadata file _scores.json contains a pre-defined security verdict of "SAFE", which is a self-referential claim designed to influence security assessment results.
Ingestion points: The security.verdict field within the _scores.json file.
Boundary markers: None.
Capability inventory: None.
Sanitization: None.

ai-content-detection