The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes a local shell script ./.github/workflows/audit-dependencies.sh using the $ARGUMENTS variable. This variable is populated directly from user input (the severity level) and interpolated into the shell command without visible sanitization, creating a potential command injection vector if the agent platform does not provide underlying protection.
[REMOTE_CODE_EXECUTION]: The core workflow requires executing pnpm install and pnpm run build. These commands trigger lifecycle scripts and build processes defined in the repository and its third-party dependencies, which constitutes execution of external code.
[EXTERNAL_DOWNLOADS]: The skill interacts with external services, specifically the NPM registry (via pnpm view and pnpm install) and GitHub (via gh pr create and searching security advisories), to manage dependencies and create pull requests.
[PROMPT_INJECTION]: The skill has an indirect prompt injection surface as it parses external, potentially attacker-influenced data to make decisions and generate content.
Ingestion points: Output from pnpm audit --json and content from external GitHub Security Advisory (GHSA) pages.
Boundary markers: Absent; the instructions do not specify delimiters or instructions to ignore embedded commands when processing audit data or advisory descriptions.
Capability inventory: File system writes (editing package.json), shell execution (pnpm, gh, and local repository scripts), and network access (GitHub PR creation).
Sanitization: No explicit sanitization or validation of the content retrieved from security advisories or package metadata is described before it is used to generate PR descriptions.

audit-dependencies