The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill implements an exec_script command that enables the execution of arbitrary JavaScript code within a browser tab. This is demonstrated in scripts/debug-chat.ts and scripts/inspect-gemini.ts, providing a capability that could be misused to manipulate web sessions or access sensitive data in the browser context.
[COMMAND_EXECUTION]: Several scripts, including scripts/deep-research.ts, scripts/list-tabs.ts, and scripts/status.ts, use Bun.spawn to execute the local command-line tools mosquitto_pub and mosquitto_sub for MQTT communication. This represents a local process execution surface.
[EXTERNAL_DOWNLOADS]: The skill requires the installation of the 'Gemini Proxy Extension' from an external, non-trusted GitHub repository (github.com/laris-co/claude-browser-proxy). Dependency on unverified third-party software components introduces potential supply chain risks.
[DATA_EXFILTRATION]: The skill provides actions to extract full HTML (get_html) and text (get_text) content from browser tabs, which are then transmitted via an MQTT broker. If the broker is not properly secured, this creates a channel for exposing sensitive chat history or page content from the Gemini interface.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection vulnerabilities. 1. Ingestion points: scripts/youtube-transcribe.ts (reads transcribed external video content). 2. Boundary markers: Absent in prompt templates. 3. Capability inventory: Arbitrary JavaScript execution via exec_script and local command execution via Bun.spawn. 4. Sanitization: No sanitization of external content was observed before processing.

gemini