The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external project files (such as package.json, requirements.txt, and lockfiles) to generate reports on vulnerabilities, licenses, and upgrades. This creates a surface for indirect prompt injection, where malicious metadata or version strings within a project's manifest could be crafted to influence the decisions of an AI agent consuming the reports.
Ingestion points: Project manifests and lockfiles are ingested by dep_scanner.py, license_checker.py, and upgrade_planner.py across various ecosystems (npm, pypi, go, cargo, rubygems).
Boundary markers: The tool generates reports in structured JSON or delimited text formats, providing basic separation between the untrusted data and the agent's instructions.
Capability inventory: The provided Python scripts are focused on file reading and report generation. No high-risk capabilities such as network exfiltration, dynamic code execution (eval/exec), or command injection via subprocess were identified in the core logic.
Sanitization: The scripts use regular expressions and specific parsing logic to extract data, which mitigates simple command injection but does not provide specific sanitization against adversarial content intended to manipulate an LLM's context.

dependency-auditor