behind-my-head
Pass
Audited by Gen Agent Trust Hub on Oct 23, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSREMOTE_CODE_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The
scripts/matte.pyscript is designed to download a pre-trained machine learning model from a third-party GitHub repository (PeterL1n/RobustVideoMatting) using thetorch.hub.loadfunctionality. - [REMOTE_CODE_EXECUTION]: In
scripts/matte.py, the skill usestorch.hub.loadwith thetrust_repo=Trueargument. This triggers the execution of the remote repository's configuration script (hubconf.py) to initialize the model on the local machine. - [INDIRECT_PROMPT_INJECTION]: The skill processes user-provided inputs such as URLs, labels, and numbers which are interpolated into a JavaScript template for video rendering.
- Ingestion points: Untrusted data enters via the
behindMyHeadfunction parameters intemplates/behind-my-head.js(e.g.,o.page,o.url,o.number,o.label). - Boundary markers: No specific boundary markers or "ignore instructions" warnings are present to encapsulate these interpolated values, though they are intended for visual rendering.
- Capability inventory: The skill environment supports executing Python scripts (
matte.py,plate.py) and Node.js rendering processes via shell commands. - Sanitization: The skill implements an
escfunction to escape basic HTML characters (&,<,>,"), which mitigates some risks of injection within the Puppeteer rendering environment.
Audit Metadata