bmad-review

Warn

Audited by Gen Agent Trust Hub on Nov 13, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill executes an external script located at {project-root}/_bmad/scripts/resolve_customization.py. Because the project-root is derived from the repository currently being worked on, a malicious repository could include a compromised version of this script. When the agent attempts to resolve configuration, it would execute the malicious script on the user's system.
  • [DYNAMIC_EXECUTION]: The skill processes activation_steps_prepend, activation_steps_append, and on_complete fields from the configuration files. These fields are treated as literal instructions for the agent to follow. Since these configurations can be overridden by project-specific files (_bmad/custom/bmad-review.toml), a malicious project could inject arbitrary instructions that the agent would then execute automatically.
  • [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze untrusted external content, such as code diffs and documentation. It lacks robust sanitization for this content, which is then passed to various "review lenses" (sub-prompts).
  • Ingestion points: The skill reads content (files, diffs), claims, and also_consider inputs from the environment.
  • Boundary markers: The instructions attempt to restrict subagents (e.g., "Do not invoke any skill"), but these are natural language constraints rather than hard technical boundaries.
  • Capability inventory: The skill has access to uv run for script execution and can read/write files in the local filesystem.
  • Sanitization: There is no evidence of filtering or escaping logic to prevent instructions embedded in the reviewed content from influencing the agent's behavior.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Nov 13, 2026, 10:20 PM