The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted codebases as input for multiple AI agents. Maliciously crafted code or comments in the target directory could potentially inject instructions into the audit agents' context, attempting to manipulate the final report or influence agent behavior. The skill attempts to mitigate this with a 'READ-ONLY' preamble, but the ingestion of arbitrary file content remains an attack surface.
[COMMAND_EXECUTION]: The workflow involves executing various command-line utilities within the target project directory to gather metadata and audit dependencies. These include tokei, find, wc, cargo audit, npm audit, pip-audit, and govulncheck. While these are standard tools, they are executed via shell which could be risky if target directory paths are not properly handled.
[CREDENTIALS_UNSAFE]: A hardcoded sensitive string is present in evals/fixture/app/config.py: API_KEY = "fake-" + "prod-api-key-do-not-use". While this is intended as a 'planted bug' for evaluation purposes, it serves as a detection point for hardcoded secrets.
[DYNAMIC_EXECUTION]: The skill uses a Workflow tool defined in references/workflow-template.md which executes a JavaScript script to coordinate the find and verify phases. This script dynamically assembles agent prompts and processes JSON outputs at runtime.

codebase-audit