The Agent Skills Directory

[PROMPT_INJECTION]: Indirect prompt injection surface detected. The skill processes untrusted source code and interpolates identifiers into shell commands without sanitization.\n
Ingestion points: SKILL.md (shell loops iterating over files and functions).\n
Boundary markers: Absent. No delimiters are used for code snippets.\n
Capability inventory: Bash tool used for file system operations and test execution.\n
Sanitization: Absent. Variables extracted from source code are used unquoted in shell loops.\n- [COMMAND_EXECUTION]: Bash loops in sections 3a and 3c (e.g., fname=$(echo "$line" | grep -oP ...) and grep -rn "$fname") use unquoted variable interpolation. This is vulnerable to command injection if a repository contains malicious file or function names designed to break shell execution.\n- [REMOTE_CODE_EXECUTION]: The skill facilitates the execution of the project's own test suites (e.g., npx jest, pytest, mvn test). Since this executes code contained within the analyzed repository, it represents a risk if the codebase being audited is malicious.\n- [EXTERNAL_DOWNLOADS]: The skill triggers the installation of tools like vulture via pip and executes packages via npx. These actions download code from public package registries.

test-coverage