motion-planning
Pass
Audited by Gen Agent Trust Hub on Nov 19, 2026
Risk Level: SAFEDYNAMIC_EXECUTIONOBFUSCATIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [DYNAMIC_EXECUTION]: The skill instructions require the agent to generate and execute Python code using shell heredocs (e.g., `python3
- << 'EOF'`) to interact with the motion planning library. This dynamic code generation is used to script robot control logic at runtime.
- [OBFUSCATION]: The instructions include a 'CRITICAL' rule forbidding the agent from using file inspection commands (
cat,head,tail) on themotion_planner.pyscript or robot XML files. This prevents the agent from inspecting or verifying the implementation details of the code it is directed to run. - [INDIRECT_PROMPT_INJECTION]: The skill ingests data from external camera images and robot configuration files (MJCF). This ingestion surface, coupled with the ability to execute robot commands, constitutes a potential vulnerability to instructions embedded in external data. Evidence:
- Ingestion points:
camera_image_pathsandmjcf_pathparameters inSKILL.md. - Boundary markers: Not present.
- Capability inventory: Subprocess execution of Python scripts and robot trajectory control.
- Sanitization: No sanitization or validation of the input files is described.
Audit Metadata