subagents-creator

Pass

Audited by Gen Agent Trust Hub on Feb 17, 2026

Risk Level: SAFENO_CODEPROMPT_INJECTION
Full Analysis
  • [Indirect Prompt Injection] (LOW): The skill's framework for creating subagents (e.g., librarian, explore) facilitates the ingestion of data from untrusted external sources, creating a vulnerability surface.\n
  • Ingestion points: As described in references/subagent-types.md, the librarian agent fetches from "docs, OSS, web" and the explore agent reads from the local codebase.\n
  • Boundary markers: The skill mandates a 7-section delegation structure (documented in SKILL.md) that includes a "MUST NOT DO" section to define forbidden actions and mitigate potential rogue behavior from processed content.\n
  • Capability inventory: Subagents are granted access to tools such as webfetch, GitHub CLI, Grep, and Read via the background_task orchestration system.\n
  • Sanitization: The skill relies on natural language constraints (the "MUST NOT DO" section) rather than programmatic sanitization or escaping of ingested data.\n- [No Code] (SAFE): The skill consists entirely of markdown files and does not contain any executable scripts, binary files, or automated installation procedures.
Audit Metadata
Risk Level
SAFE
Analyzed
Feb 17, 2026, 06:15 PM