bmad-deep-recon

Pass

Audited by Gen Agent Trust Hub on Nov 13, 2026

Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill makes extensive use of CLI tools to perform its core research functions.
  • Evidence:
  • Uses git grep, git show, git log, git blame, and git ls-tree to analyze codebase history and content across multiple files (e.g., references/codebase-analyzer.md, references/codebase-historian.md).
  • Uses gh pr list and gh pr view to retrieve pull request metadata and descriptions.
  • [DYNAMIC_EXECUTION]: Executes Python scripts using the uv run command to manage logs, configuration, and mechanical research tasks.
  • Evidence:
  • SKILL.md executes resolve_customization.py, resolve_config.py, and memlog.py located in the project's _bmad/scripts/ directory.
  • Multiple reference files and the main skill body execute the scripts/recon_kit.py helper script provided with the skill.
  • [EXTERNAL_DOWNLOADS]: Fetches data from the web for research purposes and suggests the installation of supplemental skills.
  • Evidence:
  • The 'Run' intent (described in references/run.md) performs web searches and extraction using harness-provided tools or configured MCP services.
  • SKILL.md suggests installing the bmad skill via npx skills add bmad-code-org/BMAD-METHOD if the framework is not detected.
  • [INDIRECT_PROMPT_INJECTION]: The skill possesses a significant attack surface as it is designed to ingest and process untrusted data from web search results, external research reports, and repository code.
  • Ingestion points: Processes external report files (references/process.md), web content (references/run.md), and codebase files (references/codebase.md).
  • Boundary markers: Implements a 'Research Firewall' rule that strictly forbids project context from being used as research evidence; utilizes subagents to filter raw data into 'digests' before the main agent processes them.
  • Capability inventory: Performs file system reads/writes, executes CLI commands, runs local Python scripts, and triggers web search operations.
  • Sanitization: Includes a recon_kit.py script specifically designed to strip code fences from research artifacts and escape URLs in generated HTML briefings to prevent malicious script execution.
Audit Metadata
Risk Level
SAFE
Analyzed
Nov 13, 2026, 10:20 PM