The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze untrusted third-party codebases and infrastructure configurations. To mitigate the risk of indirect prompt injection, it includes an explicit 'Untrusted input boundary' constraint in SKILL.md, instructing the agent to never interpret or execute instructions or directives found within the analyzed content.
[COMMAND_EXECUTION]: The skill suggests the use of various CLI and Docker-based security scanners. All such executions are strictly governed by a protocol requiring the agent to display the exact, pinned command and wait for explicit user approval before proceeding. It avoids arbitrary shell execution and relies on documented, verifiable command templates.
[EXTERNAL_DOWNLOADS]: The skill references established security tools such as OSV-Scanner (Google), Gitleaks, and Semgrep. These downloads are directed toward well-known, official repositories and registries. The skill enforces version pinning and integrity checks (checksums) for these external resources.
[DATA_EXFILTRATION]: While the skill accesses sensitive files (e.g., .env, IaC configs) to identify vulnerabilities, it does not exfiltrate this data. Instead, it implements a security-first storage convention using .memory/local/ and automated .gitignore management to ensure that discovered vulnerabilities and local tool outputs are not committed to version control, particularly for public repositories.
[PRIVILEGE_ESCALATION]: The skill provides instructions for users to adjust host settings (such as vm.max_map_count for SonarQube) using sudo or wsl commands. These are educational instructions for the user and are not executed autonomously or silently by the agent.

sar-cybersecurity