security-threat-model
Pass
Audited by Gen Agent Trust Hub on Sep 14, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill analyzes repository source code to produce threat models, which involves processing potentially untrusted data. Maliciously crafted content within a target repository (e.g., in code comments, documentation, or metadata) could attempt to influence the LLM's output, hide real threats, or bias the security assessment.
- Ingestion points: Codebase files located at user-provided repository paths and sub-paths as described in the
SKILL.mdworkflow andreferences/prompt-template.md. - Boundary markers: Uses a disciplined 'System prompt' and 'Evidence grounding rules' in
references/prompt-template.mdto ensure architectural claims are anchored to specific repo paths and symbol names. - Capability inventory: Requires file system read access to explore the target repository and write access to generate the final Markdown threat model report.
- Sanitization: Explicitly instructs the agent to follow 'Security hygiene' by redacting any secrets, tokens, or keys encountered in the code and verifying findings against evidence anchors.
Audit Metadata