agent-orchestrator

Pass

Audited by Gen Agent Trust Hub on Oct 11, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill establishes an orchestration pattern where a root agent processes data and code changes generated by subagents, which represents a surface for indirect prompt injection if subagents interact with untrusted content.
  • Ingestion points: The root agent ingests data from subagents in the form of Findings, Summary reports, and file diffs through the subagent tool.
  • Boundary markers: The skill mandates a structured "Delegation contract" using explicit headers (Objective, Scope, Context, Constraints, Deliverable, Acceptance criteria) to clearly delimit instructions and prevent data-instruction confusion.
  • Capability inventory: The root agent is granted capabilities for task decomposition, integration, and verification, and uses the subagent tool to spawn child processes.
  • Sanitization: The root agent is instructed to perform a mandatory "Handoff acceptance" check, comparing all subagent outputs and diffs against the original acceptance criteria and project rules before integration occurs.
  • [SAFE]: The skill design explicitly minimizes risk by defining logical roles with restricted operational scopes (e.g., the 'explorer' role is strictly read-only) and requiring the root agent to maintain oversight of all architectural and security-sensitive decisions.
Audit Metadata
Risk Level
SAFE
Analyzed
Oct 11, 2026, 08:17 PM