bad

Warn

Audited by Gen Agent Trust Hub on Apr 21, 2026

Risk Level: MEDIUMPROMPT_INJECTIONCOMMAND_EXECUTIONDATA_EXFILTRATIONEXTERNAL_DOWNLOADS
Full Analysis
  • [PROMPT_INJECTION]: The skill instructs subagents to use 'yolo mode' and 'Auto-approve all tool calls' to bypass user confirmation prompts for tool execution.
  • [COMMAND_EXECUTION]: The skill modifies the agent's configuration file ('.claude/settings.local.json') to install persistent 'PostToolUse' and 'statusLine' hooks that execute shell commands automatically.
  • [DATA_EXFILTRATION]: An installed activity log hook records tool usage details, including potentially sensitive inputs, to files located in the user's home directory ('~/.claude/projects/...').
  • [PROMPT_INJECTION]: The skill processes data from GitHub PRs and local files and passes it to autonomous subagents without using boundary markers or sanitization, creating an attack surface for indirect prompt injection.
  • Ingestion points: External GitHub PR content and local project files ('_bmad-output/planning-artifacts/epics.md', '_bmad-output/implementation-artifacts/sprint-status.yaml').
  • Boundary markers: Absent from prompts involving untrusted data.
  • Capability inventory: Use of the 'Agent' tool to spawn subagents with 'Bash' and 'FileWrite' capabilities.
  • Sanitization: No evidence of escaping or validation for external data before prompt interpolation.
  • [EXTERNAL_DOWNLOADS]: Fetches PR, issue, and status information from GitHub's API using curl for automated pipeline updates.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Apr 21, 2026, 10:04 PM