visualize-in-3d

Pass

Audited by Gen Agent Trust Hub on Nov 18, 2026

Risk Level: SAFECOMMAND_EXECUTIONREMOTE_CODE_EXECUTIONEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTIONPERSISTENCE
Full Analysis
  • [COMMAND_EXECUTION]: The skill executes several system binaries to perform its 3D modeling and video processing tasks:
  • Blender: Invoked with --python to run procedural modeling scripts.
  • ffmpeg and ffprobe: Used for video concatenation, encoding contact sheets, and auditing media streams.
  • node and npx: Used to run the author's local scripts and the Crucible MCP server.
  • playwright: Specifically used to launch headless Chromium instances for frame-by-frame recording of WebGL scenes.
  • [REMOTE_CODE_EXECUTION]: The skill includes instructions to dynamically download and execute code from the npm registry using npx --package @codepawlhq/crucible crucible-mcp. This is used to instantiate a tool-providing MCP server for the agent, which is a core part of the Crucible engine workflow.
  • [EXTERNAL_DOWNLOADS]: The skill retrieves a wide variety of assets from remote sources:
  • Hugging Face: Downloads 3D model packs (e.g., codepawl/abo-glb) and asset indices.
  • Cloud Providers: Fetches terrain and elevation data from AWS S3 buckets and USGS servers.
  • Trusted Organizations: References assets from NASA, the Smithsonian, and OpenStreetMap.
  • Safety Rule: The skill explicitly instructs the agent to "name the source, count and rough size, and get the user's OK first" before performing any downloads.
  • [INDIRECT_PROMPT_INJECTION]: The skill exposes a vulnerability surface for indirect prompt injection through data ingestion:
  • Ingestion points: Processes user-provided subjects and external data files (JSON, CSV, JSONL) for visualization (e.g., airport radar feeds, wildfire perimeters).
  • Capability inventory: The skill utilizes powerful capabilities including subprocess spawning (Blender, Python), file-write operations, and network downloads.
  • Boundary markers: No specific boundary markers are defined for untrusted data processed by the 3D rendering pipeline.
  • [PERSISTENCE]: The skill instructs the agent to persist its toolset by adding an MCP server (claude mcp add crucible-scene). This ensures that the rendering capabilities provided by the Crucible engine remain available in future sessions on the same machine.
Audit Metadata
Risk Level
SAFE
Analyzed
Nov 18, 2026, 05:07 AM