skills/edwinhu/workflows/farm-out/Gen Agent Trust Hub

farm-out

Pass

Audited by Gen Agent Trust Hub on Sep 7, 2026

Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill uses Bash scripts (farm.sh, farm-team.sh) to execute local CLI wrappers such as claude-code, codex-code, and gemini-code. These tools are used to run agent tasks in separate processes with their own execution context.\n- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted input from task files (JSON format) and interpolates this data into prompts for delegated agents.\n
  • Ingestion points: The scripts/farm.sh script reads task definitions, including prompts and labels, from a file specified via the --tasks argument.\n
  • Boundary markers: The script automatically appends an anti-simulation clause to instructions to discourage the delegated model from fabricating completion reports.\n
  • Capability inventory: Delegated processes are full agent sessions capable of tool use, filesystem modifications, and network interaction via a local proxy.\n
  • Sanitization: Content from the task files is processed using jq for structure, but the natural language prompts are not sanitized before interpolation.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 7, 2026, 05:55 AM
Security Audit — agent-trust-hub — farm-out