The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill dynamically executes third-party code using npx -y modern-web-guidance@latest search. This command downloads and runs a package from the public npm registry at runtime without version pinning or integrity checks, posing a risk of executing malicious code if the package is compromised.
[EXTERNAL_DOWNLOADS]: The skill performs several external operations: it installs the playwright package via npm if not detected, and the --publish-sample feature clones an external repository (paolomoz/stardust-site) to stage project data.
[DATA_EXFILTRATION]: Through the --publish-sample command, the skill collects local project artifacts, including rendered HTML prototypes, design tokens (DESIGN.json), and brand identity details, and uploads them to a public GitHub repository (paolomoz/stardust-site) by creating a Pull Request. While this is an intended feature for showcasing designs, it constitutes a transmission of potentially sensitive project data to an external, third-party infrastructure.
[DYNAMIC_EXECUTION]: The skill executes a local Node.js script fixtures/mobile-nav-audit.mjs which leverages Playwright to perform runtime analysis of generated prototypes. It also generates and injects complex, multi-functional JavaScript (reference/motion-runtime.md) into HTML files created by the agent.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from captured web pages via stardust/current/pages/<slug>.json (Ingestion points). There are no explicit prompt boundary markers or sanitization procedures described for this external content (Sanitization/Boundaries). Given the skill's capabilities to execute shell commands, write to the filesystem, and open network-accessible previews (Capability inventory), this creates a vulnerability where malicious instructions embedded in a target website could influence the agent's behavior during the 'craft' or 'critique' phases.

prototype