The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses subprocess.run to invoke the claude CLI when an API key is not present. The implementation is secure as it avoids shell interpolation by passing arguments as a list and provides data via stdin rather than command-line arguments.
[DATA_EXFILTRATION]: The skill transmits file contents to the Anthropic API for compression. It implements a rigorous heuristic check (is_sensitive_path) to identify and block the processing of files containing secrets, such as environment variables, SSH keys, or credential stores, ensuring sensitive data is not exfiltrated to the third-party API.
[INDIRECT_PROMPT_INJECTION]: As the skill processes natural language files through an LLM, it contains a surface for indirect prompt injection attacks.
Ingestion points: Untrusted content is read from user-provided file paths in scripts/compress.py via the read_source function.
Boundary markers: The compression prompt identifies the source text with a "TEXT:" label but lacks strict structural delimiters to isolate the untrusted content.
Capability inventory: The skill can read and write files on the local filesystem and execute the claude binary.
Sanitization: The skill mitigates potential injection effects using a validation engine in scripts/validate.py that confirms structural integrity (e.g., ensuring no code blocks or URLs were altered) before accepting the output.

caveman-compress