The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes game code and visits URLs using Playwright, which creates a surface for indirect prompt injection attacks. A malicious game or website could attempt to influence the agent's behavior through manipulated console logs, visual content in screenshots, or structured state data. \n
Ingestion points: Game source code, game URL, and the output of window.render_game_to_text() (referenced in SKILL.md and scripts/web_game_playwright_client.js). \n
Boundary markers: The skill does not explicitly define delimiters or warnings to ignore instructions embedded within the game data. \n
Capability inventory: The agent can execute shell commands (node), write to the filesystem (fs.writeFileSync in scripts/web_game_playwright_client.js), and perform arbitrary network navigation via Playwright. \n
Sanitization: No sanitization or validation of the game state or console output is mentioned before being processed by the agent.\n- [EXTERNAL_DOWNLOADS]: The skill instructions in SKILL.md suggest installing the @playwright/mcp@latest package globally via npm if Playwright is not available. While originating from a well-known scope, this is an external dependency installation.\n- [COMMAND_EXECUTION]: The workflow requires the agent to run a local Node.js script (scripts/web_game_playwright_client.js) to perform game testing. This script executes automated browser interactions and writes output files (screenshots and state JSON) to the local directory output/web-game.\n- [DYNAMIC_EXECUTION]: The Playwright client uses page.evaluate() and page.addInitScript() to inject and execute JavaScript within the browser context of the game being tested (found in scripts/web_game_playwright_client.js).\n- [REMOTE_CODE_EXECUTION]: The Playwright script navigates to a user-provided or agent-provided URL and executes the associated code within the browser context, which represents a managed form of remote code execution.

develop-web-game