scenario-game-ua-ads

Pass

Audited by Gen Agent Trust Hub on Nov 9, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONDATA_EXFILTRATION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill ingests and processes untrusted user data which could contain malicious instructions.
  • Ingestion points: The skill instructions in SKILL.md direct the agent to process a "Concept or the game's core loop", "game name", and "game's own material" including screenshots and video recordings.
  • Boundary markers: The instructions lack explicit delimitation (such as XML tags or unique markers) for user-supplied data and do not provide a clear "ignore embedded instructions" directive to the agent when processing this content.
  • Capability inventory: The skill possesses significant capabilities via bundled Python scripts (render_tracks.py, build_passes.py, check_master.py, check_take.py, put_parts.py, make_patch.py) that can read/write files, execute shell commands, and perform network operations.
  • Sanitization: While scripts like put_parts.py validate network destinations, there is no high-level sanitization of user-provided text or media metadata before it is used to influence agent behavior or script arguments.
  • [COMMAND_EXECUTION]: Multiple scripts execute local shell commands to facilitate media production tasks.
  • Evidence: The scripts render_tracks.py, build_passes.py, check_master.py, check_take.py, and make_patch.py utilize subprocess.run() to invoke the ffmpeg and ffprobe binaries.
  • The arguments for these commands are dynamically generated based on JSON specifications (plan.json, spec.json) produced by the agent. This creates a risk of command argument injection if the agent fails to properly sanitize inputs derived from user data.
  • [DATA_EXFILTRATION]: The skill includes a utility for uploading files to external storage.
  • Evidence: scripts/put_parts.py uses the http.client library to perform HTTP PUT operations to upload local files to remote URLs.
  • The script implements a security check to ensure that the destination host ends with .amazonaws.com (Amazon S3) or is a loopback address. While this aligns with the functionality of the Scenario platform, it represents a mechanism for sending local data to an external well-known service.
Audit Metadata
Risk Level
SAFE
Analyzed
Nov 9, 2026, 12:14 AM