The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted text from construction contracts, creating a surface where malicious instructions could be embedded in documents to influence agent behavior.
Ingestion points: The ContractClauseExtractor.extract_clauses and generate_summary methods in SKILL.md process user-provided contract_text strings.
Boundary markers: The implementation lacks explicit delimiters or instructions to ignore embedded commands within the processed text.
Capability inventory: The skill requests filesystem permissions in claw.json, which allows the agent to perform file operations that could be targeted by an injection attack.
Sanitization: The code uses regex for data extraction and truncates text to 2000 characters, but does not sanitize the content against instructions directed at the LLM.
[SAFE]: The skill implementation uses standard Python libraries (re, dataclasses, enum) to perform local text analysis. No network requests, external downloads, or obfuscation techniques were detected. The homepage and author identifiers are consistent with the vendor's own domain.

contract-clause-extractor