The Agent Skills Directory

[PROMPT_INJECTION]: The skill implements detection logic to identify potential prompt injection attempts within the audited project. It flags phrases like 'ignore previous instructions' to warn the agent and user of hostile content. Reference files in the skill itself mention these phrases for documentation of security practices and do not constitute an attempt to bypass agent guardrails.
[CREDENTIALS_UNSAFE]: The skill includes heuristic patterns to detect sensitive information such as AWS keys, GitHub tokens, and private keys in the target project's documentation. This feature is purely defensive and serves to prevent the accidental commitment of secrets to version control.
[COMMAND_EXECUTION]: The audit logic utilizes git ls-files to check the status of specific local files (e.g., SCRATCHPAD.md). This command is hardcoded with specific filenames and does not accept arbitrary user input, preventing shell injection vulnerabilities.
[INDIRECT_PROMPT_INJECTION]: The skill audits untrusted markdown files from the local project. It manages this risk by serving as a detection tool for malicious instructions and by requiring explicit user confirmation before performing any write operations. Ingestion occurs via fs.readFileSync in scripts/audit.mjs, and capabilities are limited to filesystem writes (fs.writeFileSync) and version control status checks.

memory-audit