subagent-driven-development

Pass

Audited by Gen Agent Trust Hub on Apr 4, 2026

Risk Level: SAFEPROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
  • [PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it ingests implementation plans and task metadata from external sources (e.g., plan files in docs/superpowers/plans/ and task descriptions via TaskGet). Malicious instructions embedded within these plans could attempt to override the subagents' behavior.\n
  • Ingestion points: Plan files (.md), task descriptions, and json:metadata code fences within tasks.\n
  • Boundary markers: Uses markdown headers (e.g., ## Task Description) and JSON structures to delineate data, but these do not provide robust security isolation against adversarial content.\n
  • Capability inventory: The subagents can modify the filesystem, commit to git, and execute shell commands.\n
  • Sanitization: The skill lacks explicit sanitization or validation logic for the content of the plans it processes.\n- [COMMAND_EXECUTION]: The implementer subagent is instructed to execute a verifyCommand provided in the task metadata. This allows for the execution of arbitrary shell commands. While this is a functional requirement for automated testing (TDD), it poses a security risk if the verification command is sourced from an untrusted or compromised implementation plan.
Audit Metadata
Risk Level
SAFE
Analyzed
Apr 4, 2026, 10:10 PM