subagent-driven-development

Pass

Audited by Gen Agent Trust Hub on Mar 14, 2026

Risk Level: SAFEPROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
  • [PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface by processing data from an external system (beads) while possessing significant capabilities.\n
  • Ingestion points: Subagents execute bd show and bd comments (detailed in context-loading.md and implementer-prompt.md) to retrieve requirements and wave conventions directly from issue descriptions and comments.\n
  • Boundary markers: While prompt templates in implementer-prompt.md use structural headers, they do not employ unique delimiters or specific instructions to disregard embedded commands in the retrieved data.\n
  • Capability inventory: Agents can modify the filesystem (Write tool), execute shell commands via git and npm, and manipulate issue states (detailed in background-execution.md and implementer-prompt.md).\n
  • Sanitization: There is no evidence of filtering or sanitizing the content retrieved from the beads system before it is incorporated into the agent's context.\n- [COMMAND_EXECUTION]: The skill relies on extensive execution of system commands and CLI tools to perform its orchestration duties.\n
  • Evidence: Files such as background-execution.md, implementer-prompt.md, and verification-and-evidence.md specify the use of git for version control, npm for testing and building, and the bd utility for managing the beads environment.
Audit Metadata
Risk Level
SAFE
Analyzed
Mar 14, 2026, 09:33 AM