cso
Pass
Audited by Gen Agent Trust Hub on Oct 27, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and audit untrusted source code, Git history, and other repository metadata, which creates a significant surface for indirect prompt injection.
- Ingestion points: The skill explicitly reads source code, Git history, and external skill definitions (SKILL.md) through various helper commands (SKILL.md, Phases 0, 2, and 8).
- Boundary markers: The skill instructions emphasize treating repository content as untrusted evidence and utilizing a trusted helper for redaction and containment (SKILL.md, Phase 0).
- Capability inventory: The skill uses a launcher binary to execute security scans (Gitleaks, Semgrep, Trivy), manage Docker containers, and write findings to the filesystem (SKILL.md, Phases 12-13).
- Sanitization: The skill implements strict environment hygiene rules to prevent the accidental disclosure of host environment variables during execution (SKILL.md).
- [COMMAND_EXECUTION]: The skill requires execution of a specific local binary to perform its auditing and scanning operations.
- Evidence: The allowed-tools section in SKILL.md and SKILL.md.tmpl permits the execution of ~/.claude/skills/gstack/bin/gstack-cso-launcher and its Windows executable equivalent.
Audit Metadata