gathering-feature
Pass
Audited by Gen Agent Trust Hub on Oct 9, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted feature specifications which are used to orchestrate a chain of subagents, creating a surface for potential instruction injection through data.
- Ingestion points: The conductor skill accepts a "feature request" from the user (SKILL.md, Phase 1), which is then passed as the primary mission to the Bloodhound scout and subsequent agents.
- Boundary markers: The skill employs strong isolation techniques by using structured handoff templates (references/conductor-dispatch.md) that limit the context provided to each subagent; for example, the Turtle hardening agent receives only a file list rather than the full implementation history.
- Capability inventory: The orchestrated subagents have capabilities to read/write files in the repository and execute shell commands using project tools (gw, gf) and standard package managers (pnpm, uv).
- Sanitization: The skill mandates strict sanitization practices including Zod schemas for all entry points, context-aware output encoding, and Rootwork utilities for type safety at trust boundaries. It also includes a dedicated adversarial hardening phase (Turtle) and a parallel security audit (Raccoon) to identify and fix vulnerabilities.
Audit Metadata