orchestrating-swarms

Pass

Audited by Gen Agent Trust Hub on Apr 10, 2026

Risk Level: SAFEEXTERNAL_DOWNLOADSPROMPT_INJECTION
Full Analysis
  • [SAFE]: The skill is primarily instructional documentation for a platform's orchestration system. All code examples and logic represent legitimate use of the TeammateTool and Task system for multi-agent coordination.\n- [EXTERNAL_DOWNLOADS]: The documentation references installing the it2 CLI tool via standard package managers (uv, pipx, pip) to enable visible panes in iTerm2. This is a well-known utility for terminal integration and is documented neutrally as environment configuration.\n- [PROMPT_INJECTION]: The orchestration architecture facilitates indirect prompt injection by allowing subagents to ingest tasks and messages from shared files. This is a documented vulnerability surface inherent to the multi-agent design described.\n
  • Ingestion points: Subagents read from files in ~/.claude/tasks/ and ~/.claude/teams/*/inboxes/.\n
  • Boundary markers: Example prompts interpolate data directly without specific delimiters or warnings.\n
  • Capability inventory: Teammates can be spawned with the general-purpose type, which has access to all agent tools.\n
  • Sanitization: No input validation or sanitization mechanisms are specified in the provided orchestration patterns.
Audit Metadata
Risk Level
SAFE
Analyzed
Apr 10, 2026, 02:13 PM