caveman-compress

Pass

Audited by Gen Agent Trust Hub on Jul 10, 2026

Risk Level: SAFECOMMAND_EXECUTIONDATA_EXFILTRATIONPROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill executes the claude CLI tool using subprocess.run() in scripts/compress.py when an API key is not configured. This is a functional requirement to perform the compression. The implementation uses a list-based argument format without a shell, which prevents command injection.\n- [DATA_EXFILTRATION]: The skill transmits file content to the Anthropic API for processing. It includes a proactive security filter (is_sensitive_path in scripts/compress.py) that uses regex and path analysis to detect and refuse files likely containing secrets, private keys, or credentials (e.g., .env, .ssh/, id_rsa) before they are sent to the external API.\n- [PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection as it processes untrusted user files and interpolates their content into LLM prompts. This is mitigated by a validation engine in scripts/validate.py that verifies the structural integrity of the compressed output (headings, code blocks, URLs) against the original file, reverting changes if discrepancies are detected.\n
  • Ingestion points: Reads the content of user-provided file paths in scripts/compress.py.\n
  • Boundary markers: The prompts in build_compress_prompt use explicit instruction blocks and clear labeling of the input text.\n
  • Capability inventory: The skill has permissions to read and write to the target file and communicate with the Anthropic API.\n
  • Sanitization: Relies on structural validation of the output rather than input sanitization.
Audit Metadata
Risk Level
SAFE
Analyzed
Jul 10, 2026, 12:22 PM
Security Audit — agent-trust-hub — caveman-compress