dcl-prototype
Pass
Audited by Gen Agent Trust Hub on Nov 3, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADSCREDENTIALS_UNSAFE
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill ingests data from user-authored hypothesis files which could contain instructions that influence agent behavior.
- Ingestion points: The agent reads markdown files from the
design/directory, specificallydesign/*/H*_*.mdanddesign/gdd.md. - Boundary markers: The skill relies on structured headers (e.g.,
## Brief,## Sessions) to parse content but does not include explicit instructions to disregard embedded commands or use sanitization. - Capability inventory: The skill has access to file system writes, shell command execution via
npx, and local network probing of the scene preview port. - Sanitization: No methods for escaping or filtering instructions within the ingested data are mentioned.
- [COMMAND_EXECUTION]: The skill instructs the agent to execute shell commands to manage its environment and tools.
- Evidence:
references/greybox-build.mdandreferences/bootstrap.mddescribe usingnpx skills add decentraland/sdk-skills --allandskills update. - Mitigation: The instructions contain security constraints requiring the agent to "ask the owner before installing" or launching the preview server.
- [EXTERNAL_DOWNLOADS]: The skill references several external repositories for game assets and models.
- Evidence:
references/scene-resources.mdprovides links tokenney.nl,quaternius.com,polyhaven.com,ambientcg.com, andfreesound.org. - Note: These sources are well-known and reputable services within the game development community.
- [CREDENTIALS_UNSAFE]: The skill includes guidance to avoid exposing sensitive information during its operation.
- Evidence:
references/bootstrap.mdexplicitly states: "Keep tokens, credentials and copied MCP configs out of this note."
Audit Metadata