orchestrating-subagents

Pass

Audited by Gen Agent Trust Hub on Aug 19, 2026

Risk Level: SAFE
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill documents patterns for subagents to interact with external data such as file contents and codebase search results. This represents a known attack surface for indirect prompt injection. The templates provided include boundary markers and validation rules to minimize risk.\n
  • Ingestion points: File reading and codebase search tools in SKILL.md and references/prompt-templates.md.\n
  • Boundary markers: Templates use explicit JSON formatting instructions and unique identifiers to mark completions.\n
  • Capability inventory: Tools for file reading, writing, and subagent spawning are described.\n
  • Sanitization: Instructions recommend validating outputs and tightening prompt specifications.\n- [COMMAND_EXECUTION]: Examples in references/learning-log.md demonstrate the use of the grep utility to search a local performance log file (~/.claude/subagent-runs.jsonl). These commands are restricted to local metadata and serve a diagnostic purpose.\n- [PRIVILEGE_ESCALATION]: Reference documentation in references/agent-definition.md describes permissionMode options such as bypassPermissions. These are internal platform settings for governing user prompt frequency and do not grant elevated system permissions.
Audit Metadata
Risk Level
SAFE
Analyzed
Aug 19, 2026, 06:35 PM
Security Audit — agent-trust-hub — orchestrating-subagents