The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes local commands including node and browser-harness-js. It securely manages user-supplied queries by using JSON.stringify in a Node.js wrapper to escape strings before they are injected into JavaScript templates for browser execution. This prevents command injection vulnerabilities. Additionally, the setup script modifies shell configuration files (.bashrc, .zshrc, etc.) to add the user's local bin directory to the PATH, which is a standard procedure for installing command-line tools.
[PROMPT_INJECTION]: The skill exposes an indirect prompt injection surface because it processes untrusted content from the internet.
Ingestion points: Data enters the context from google.com (via the search results tab) and is processed in the scripts/gnews script.
Boundary markers: The skill returns news results as formatted text or JSON without specific delimiters or warnings instructing the agent to ignore embedded commands within the snippets.
Capability inventory: The skill allows for local command execution and browser automation through the browser-harness-js SDK.
Sanitization: Input queries are sanitized via JSON.stringify and the browser extraction uses innerText to remove HTML formatting, but the textual content of news snippets is not filtered for potential adversarial instructions.

gnews