The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill workflow for diagnosing failures and verifying fixes involves processing untrusted LLM outputs, which can lead to indirect prompt injection vulnerabilities.
Ingestion points: The debug_llm_output and diagnose_failure functions in SKILL.md take prompt and output as arguments, which represent untrusted external data.
Boundary markers: The prompt templates generated by suggest_prompt_fix lack specific delimiters or "ignore embedded instructions" warnings to isolate potentially malicious data.
Capability inventory: The debug_llm_output function in SKILL.md invokes the llm() tool to test generated fixes, allowing injected data to influence subsequent agent actions.
Sanitization: There is no evidence of sanitization or filtering of the output or prompt data before it is interpolated into new instructions or test cases.

llm-debugger