llm-mechanical-designer
Pass
Audited by Gen Agent Trust Hub on Nov 7, 2026
Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill's test suite utilizes the subprocess module to run regression tests on the internal design loop script.\n
- Evidence: In
tests/test_truss.py,subprocess.runis used to executescripts/design_loop.pyusing the current Python interpreter.\n - Assessment: This behavior is benign as it is confined to the test environment and uses hardcoded local paths rather than untrusted user input to construct commands.\n- [INDIRECT_PROMPT_INJECTION]: The skill follows an 'LLM-in-the-loop' architecture where it ingests design data and task configurations that are subsequently used to generate prompts for the agent.\n
- Ingestion points:
scripts/design_loop.pyreads data fromdesign.json(agent output) andtask.json(user-provided or preset configuration).\n - Boundary markers: Prompts defined in
references/prompts.mdutilize structured JSON blocks and markdown headings to separate data from instructions.\n - Capability inventory: The skill performs local file operations, such as writing iteration history and updated prompt files to a specified run directory.\n
- Sanitization: Inputs are parsed using standard JSON libraries, which mitigates structural injection, though the content is directly interpolated into natural language prompts for the optimization cycle.\n
- Assessment: The surface area for exploitation is low and inherent to the intended iterative optimization functionality of the mechanical designer skill.
Audit Metadata