The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill presents an inherent indirect prompt injection surface as it is designed to ingest and analyze untrusted third-party data such as source code, infrastructure configurations, and log files.
Ingestion points: Processes project files, source code (Python, JS/TS, Go), CI/CD configurations, and external documentation at runtime (references/sast-tools.md, references/vulnerability-patterns.md).
Boundary markers: The skill relies on process-based boundaries (e.g., 'rules of engagement', 'scope verification') but lacks technical prompt boundaries or delimiters to prevent the agent from obeying instructions embedded within the analyzed data.
Capability inventory: The agent has permissions to read files and is instructed to execute various CLI-based security scanners and network tools.
Sanitization: No explicit sanitization logic or content filtering is provided for the data being analyzed.
[COMMAND_EXECUTION]: The workflow involves the execution of powerful CLI tools for security analysis, including static analysis tools (Semgrep, Bandit), network scanners (nmap), and exploitation validation tools (sqlmap). These commands are standard for security auditing and are executed at the agent's discretion during the audit process.
[EXTERNAL_DOWNLOADS]: The reference materials suggest the installation of numerous security packages and utilities from well-known package registries (npm, pip, brew, go) to facilitate the auditing process. These are legitimate tools used in the professional security industry.

security-reviewer