The Agent Skills Directory

[PROMPT_INJECTION]: The skill includes a robust defensive rule (Rule 6) that explicitly instructs the agent to ignore and report any malicious commands, such as 'ignore previous instructions', encountered within the audited codebase. This acts as a safeguard against indirect prompt injection, and the static analysis flag for this pattern is a false positive triggered by these defensive instructions.
[CREDENTIALS_UNSAFE]: Rule 4 establishes a strict policy against reproducing discovered secrets. It mandates that findings reference only the file location and credential type, ensuring that sensitive data like API keys or tokens are never included in the skill's output, plans, or GitHub issues.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted data from repositories. It mitigates this risk through mandatory evidence chains for findings and specific instructions to treat all repository content as data rather than instructions, alongside using isolated git worktrees for execution-related tasks.
[COMMAND_EXECUTION]: The skill utilizes standard development utilities (git, npm, tsc, gh) for codebase analysis. It contains strict rules against mutating the user's primary working tree and requires the use of isolated environments for any code modification tasks.
[DATA_EXPOSURE]: The skill includes a proactive security check for its GitHub issue creation feature, requiring the agent to warn the user and get explicit confirmation before publishing plans to public repositories if they contain sensitive findings.

improve