The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill's primary function is to ingest and process untrusted data from an audited project, including source code, commit history, and pull request content.
Ingestion points: Project files are read in Phase 1, and external data from the GitHub API (security alerts) is ingested in Phase 2.
Boundary markers: The skill contains an explicit 'Safety and Instruction Boundary' section that instructs the agent to treat all audited content as untrusted and to disregard any embedded instructions or requests.
Capability inventory: The skill utilizes local bash scripts, version control tools (git), and the GitHub CLI (gh) which can perform network operations and modify repository state (e.g., dismissing alerts).
Sanitization: The instructions emphasize static review before any execution and the use of disposable test environments to mitigate the risk of malicious code within the audited material.
[COMMAND_EXECUTION]: The skill facilitates the execution of a bundled discovery script and various system tools.
Evidence: The skill executes scripts/security-surface.sh, a bash script that uses ripgrep (rg) to identify security-sensitive code patterns.
Capability: It also instructs the agent to use git for repository inspection and gh api for managing security alerts, including PATCH requests that modify the state of alerts on GitHub.

security-audit