Most static backtick alerts are false positives from Markdown formatting, input tables, and sample code blocks. The confirmed issues are the documented belt and npx command flows plus the intentional dependency on hosted inference.sh rendering. No prompt injection text was found in SKILL.md.
> **Install the belt CLI skill:** `npx skills add belt-sh/cli`
The instruction asks users to execute npx to add a third-party CLI skill. This is a real supply-chain risk because it installs executable tooling from outside the current package.
The line directs users to install and use the belt CLI before running the skill. The executable command is not embedded on this line, but the dependency creates real external-command exposure.
The fenced quick-start block contains belt login and belt app run commands. Running those commands executes an external CLI and submits render input to a hosted service.
This bash example invokes belt app run with user-provided Remotion code. It is legitimate skill behavior, but it is still external CLI execution with remote processing.
This bash example invokes belt app run for a render job. The command is an intentional external service call and should remain a confirmed medium-risk finding.
The location starts another bash example that runs belt app run. The command path is expected, but it executes third-party CLI tooling and sends input outside the local environment.
This sequence animation example uses belt app run to submit TSX code for rendering. That is expected functionality, but it is still confirmed external CLI execution.
The related-skills block gives npx skills add commands for additional packages. These commands can install third-party skill content, so the supply-chain risk is real.
The Markdown image uses a hosted cloud.inference.sh asset. Viewing the documentation can trigger a network fetch to a third-party host, which is a low privacy risk.
The install-instructions link points to raw GitHub content for the required CLI. This is an intentional dependency path, but users rely on external network content to set up the tool.
The skill sends React and Remotion TSX, props, and render settings to inference.sh for hosted rendering. Users could expose proprietary animation code, unreleased content, or sensitive asset references if they submit private project data.
The documentation states rendering is done through inference.sh and shows code submitted through belt app run input. The behavior is intentional, but the data exposure risk is direct.
04
是正措置
この監査で推奨される修正が記録されています。これらを適用する責任はメンテナーにあります。
FIX-001
中
External CLI commands are required for normal use.
Require explicit user confirmation before running belt or npx commands, and document the exact commands and permissions needed.
FIX-002
中
Render inputs are sent to a hosted inference.sh service.
Warn users not to submit secrets, private code, or confidential media. Link to data handling and retention terms near the quick start.
FIX-003
低
Setup instructions depend on third-party installation sources.
Pin recommended CLI and related-skill versions when possible, and provide checksum or provenance guidance for trusted installation.