orchestrator

Warn

Audited by Gen Agent Trust Hub on Mar 15, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill is designed to execute arbitrary commands on the host system by spawning sub-processes for various LLM CLI tools via wrapper scripts like scripts/spawn-agent.sh and scripts/parallel-run.sh.- [COMMAND_EXECUTION]: In config/cli-config.yaml, the skill explicitly configures multiple LLM providers (Gemini, Claude, Codex, Qwen) to operate in 'yolo', 'dangerously-skip-permissions', or 'full-auto' modes. These flags intentionally bypass the standard user approval process for actions like file system modification or network access.- [PROMPT_INJECTION]: The resources/subagent-prompt-template.md file interpolates {TASK_DESCRIPTION} and {ACCEPTANCE_CRITERIA} directly into the prompt without sanitization, creating a surface for direct instruction override if the task source is manipulated.- [PROMPT_INJECTION]: The skill possesses a significant indirect prompt injection surface:
  • Ingestion points: The orchestrator reads untrusted data from task-board.md, progress-{agent-id}.md, and result-{agent-id}.md.
  • Boundary markers: No delimiters or 'ignore embedded instructions' warnings are present in the prompt templates when reading these shared memory files.
  • Capability inventory: The orchestrator can spawn processes, write files, and execute parallel agents via the oh-my-ag tool.
  • Sanitization: No evidence of validation, escaping, or filtering of external content exists before processing agent-generated results or updating session metrics.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Mar 15, 2026, 10:44 PM