research-implement-feature

Pass

Audited by Gen Agent Trust Hub on Nov 12, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONDYNAMIC_EXECUTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest untrusted data from user arguments and external sources (such as paper PDFs, READMEs, or issue threads) to drive the implementation of features. While the skill includes instructions to treat external content as data and references an 'injection-hygiene.md' policy, the use of this data to generate code that is then executed creates a significant attack surface.
  • Ingestion points: $ARGUMENTS in SKILL.md and external content fetched during the 'Reuse survey' or Phase 0.
  • Boundary markers: The skill explicitly instructs the agent to treat fetched content as data rather than instructions, but lacks strict schema-based validation for the generated code.
  • Capability inventory: Access to Bash(*) for unrestricted shell commands, along with Write and Edit tools for modifying the file system.
  • Sanitization: Relies on model-level instruction following to prevent external data from hijacking the implementation logic.
  • [COMMAND_EXECUTION]: The skill makes extensive use of the Bash(*) tool to autonomously execute 'acceptance checks' and the 'F0 spine' (walking skeleton) of the code it generates. In ASK=never mode, these commands are executed without user confirmation.
  • [DYNAMIC_EXECUTION]: The skill's primary workflow involves generating executable scripts (e.g., Python, shell) and running them at runtime to verify implementation rungs. This 'build and test' loop is autonomous and based on potentially untrusted input requirements.
Audit Metadata
Risk Level
SAFE
Analyzed
Nov 12, 2026, 06:07 AM