The Agent Skills Directory

[SAFE]: The skill is designed with a strong security-first approach for its primary use case of running dynamic game content.
[REMOTE_CODE_EXECUTION]: The skill implements dynamic HTML/JS execution using the srcDoc attribute in an iframe. This is secured using the sandbox="allow-scripts" attribute without the allow-same-origin flag, ensuring the execution environment is strictly isolated from the parent application's data and DOM.
[DATA_EXFILTRATION]: All communication between the parent environment and the sandboxed iframe is restricted to a private MessageChannel port transferred during initialization. This prevents message spoofing or interception that could occur with a global postMessage listener.
[PROMPT_INJECTION]: The skill handles untrusted AI-generated content (gameHtml) but mitigates indirect injection risks through a clear evidence chain: Ingestion occurs in components/GameIframe.tsx; the technical boundary is enforced by the iframe sandbox; capabilities are restricted to an answer submission flow; and all data is sanitized/validated server-side in convex/answers.ts using the checkAnswer function.

game-engine