flow-next-resolve-pr

Warn

Audited by Gen Agent Trust Hub on Apr 25, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: Phase 6 of the workflow executes a command (PROJECT_TEST_CMD) that is dynamically retrieved from repository files such as AGENTS.md or CLAUDE.md. If an attacker compromises these files, they can achieve arbitrary command execution during the validation phase.\n
  • Evidence: In workflow.md Phase 6, the skill reads the preferred command from project documentation and executes it: PROJECT_TEST_CMD="(...)" followed by $PROJECT_TEST_CMD.\n- [PROMPT_INJECTION]: The skill ingests untrusted PR comments and review data from the GitHub API, creating a surface for Indirect Prompt Injection that could manipulate the sub-agents dispatched to resolve feedback.\n
  • Ingestion points: Untrusted feedback is fetched from the GitHub API via the get-pr-comments script in workflow.md (Phase 1).\n
  • Boundary markers: The skill mentions quoting feedback (e.g., using >-prefixed lines) but lacks formal or cryptographic delimiters to distinguish untrusted data from agent instructions.\n
  • Capability inventory: The skill possesses significant capabilities, including file modification, shell command execution (via PROJECT_TEST_CMD), and git operations (commit and push).\n
  • Sanitization: The workflow includes a "Forbidden" instruction to ignore shell commands in comment bodies, but it lacks programmatic sanitization or validation of the ingested comment text before passing it to sub-agents.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Apr 25, 2026, 02:41 AM