orchestrate-subagents

Pass

Audited by Gen Agent Trust Hub on Sep 5, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill design facilitates the processing and redistribution of untrusted user-supplied objectives to subagents, presenting an indirect prompt injection attack surface.
  • Ingestion points: User-defined objectives, task descriptions, and constraints are interpolated into structured "prompt packets" (defined in references/prompt-packets.md) for subagents.
  • Boundary markers: The templates for subagent delegation do not include robust delimiters or specific instructions to isolate interpolated user data from the subagent's operational instructions.
  • Capability inventory: The skill orchestrates subagents with the capability to perform codebase modifications (references/delegation-patterns.md, Pattern E) and interact with various agent harnesses like Codex and Claude Code (references/harness-adapters.md).
  • Sanitization: There are no defined mechanisms within the orchestration workflow to validate, escape, or sanitize external content before it is processed by the subagent wave.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 5, 2026, 05:37 AM
Security Audit — agent-trust-hub — orchestrate-subagents