The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill operates by reading the entire contents of a codebase in Phase 3 ("Read every line"). This constitutes a significant ingestion point for untrusted data. If the files being analyzed contain malicious instructions (e.g., hidden in comments or string literals), they could influence the agent to deviate from the skill's safety rules or perform unauthorized deletions. The skill lacks explicit boundary markers or sanitization protocols for the ingested content. Its capabilities include deleting files, modifying code, and executing shell commands, which increases the potential impact of a successful injection.
[COMMAND_EXECUTION]: The detection cookbook (references/detection.md) recommends executing a wide variety of shell-based commands and CLI tools using npx, pip, go, and cargo. This includes commands for calculating line counts, checking for code clones, detecting unused exports, and running test suites. While these are standard development tasks, they involve the execution of external binaries and the potential for argument injection if file paths or scope names are not handled carefully.
[EXTERNAL_DOWNLOADS]: The skill's instructions facilitate the use of external tools from public registries like NPM and PyPI via npx one-shots and pip commands. While these target well-known development utilities (e.g., cloc, jscpd, knip, ruff), this behavior involves downloading and running code from external sources during the analysis phase.

nuclear-simplification