The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes content from arbitrary websites which may contain instructions targeting the AI agent. Ingestion points include page content retrieved via getAISnapshot and textContent methods in src/client.ts. The skill does not currently implement explicit boundary markers or sanitization to protect the agent from embedded instructions in the web data it processes. The agent has significant capabilities including filesystem access and shell execution which increases the potential impact of such injections.
[COMMAND_EXECUTION]: The setup process involves host-level command execution to manage browser processes. The scripts/start-server.ts script uses execSync to run lsof and kill -9 on port 9223 to ensure a clean startup environment for the Playwright server.
[DYNAMIC_EXECUTION]: The skill employs dynamic script evaluation to facilitate browser interactions. In src/client.ts, the connect function uses eval() within page.evaluate() to inject a large utility script (src/snapshot/browser-script.ts) into the browser context for generating AI-ready page snapshots.
[EXTERNAL_DOWNLOADS]: The skill automates the download of dependencies and browser binaries. The server.sh script executes npm install to fetch Node.js packages, and scripts/start-server.ts runs npx playwright install chromium to obtain the necessary browser engine. These operations target official package registries and Playwright repositories.

dev-browser