effectcraft-cli-export
Pass
Audited by Gen Agent Trust Hub on Oct 13, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADSDYNAMIC_EXECUTIONREMOTE_CODE_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied rendering projects (.ecproj) and execution plans (PLAN.json), which serves as a vector for untrusted data to influence agent behavior.
- Ingestion points: The agent ingests project state data and structured command sequences from local or external files.
- Capability Inventory: The skill has the ability to invoke shell commands, perform network downloads for assets, and write arbitrary render outputs to the filesystem.
- Boundary markers: The skill utilizes structured JSON schemas, duplicate key rejection, and pre-execution validation (check mode).
- Sanitization: It implements 'engineering verification' steps that compare project hashes before and after execution to detect unauthorized changes.
- [EXTERNAL_DOWNLOADS]: The skill installs the EffectCraft CLI and associated model weights (e.g., Roto Brush and Face Tracking models) from remote sources.
- Security Controls: All binaries, archives, and model weights are verified against SHA256 integrity hashes provided in the skill's metadata (runtime.lock.json, bridge-tools.json).
- [COMMAND_EXECUTION]: The skill manages rendering tasks by executing shell scripts (launch.sh, launch.ps1) and specialized Python wrappers.
- These executions are scoped to a managed directory (HARNESS_DIR) and use isolated Python environments to prevent unintended modification of the global system state.
- [DYNAMIC_EXECUTION]: The skill provides a 'run_script' tool that allows for the execution of arbitrary JavaScript within the rendering engine's scripting environment.
- While this is a documented feature for automating complex graphical tasks, it represents a dynamic execution surface that processes input code.
- [REMOTE_CODE_EXECUTION]: The skill initiates the installation of its own CLI components via a bootstrap script (bootstrap.py).
- These scripts originate from the vendor's own harness skill and are part of the intended environment setup, though they represent a remote code execution pattern.
Audit Metadata