skills/garrytan/gstack/pair-agent/Gen Agent Trust Hub

pair-agent

Pass

Audited by Gen Agent Trust Hub on Nov 16, 2026

Risk Level: SAFEEXTERNAL_DOWNLOADSREMOTE_CODE_EXECUTIONCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [EXTERNAL_DOWNLOADS]: The skill downloads an installation script for the Bun runtime from https://bun.sh/install. This is a well-known service, and the skill performs a mandatory integrity check using a hardcoded SHA-256 checksum (bab8acfb046aac8c72407bdcce903957665d655d7acaa3e11c7c4616beae68dd) before proceeding with any execution.
  • [REMOTE_CODE_EXECUTION]: After the checksum verification successfully validates the downloaded Bun installer, the script is executed using the system shell. This represents a secure implementation of dynamic dependency management.
  • [INDIRECT_PROMPT_INJECTION]: By enabling remote agents to browse the web through the local browser, the skill introduces a surface for indirect prompt injection. Malicious data or instructions found on web pages could potentially influence the behavior of the paired agent. The skill provides built-in mitigations via the --restrict flag, which allows users to enforce read-only access or disable JavaScript and storage operations for the remote session.
  • Ingestion points: Web page content and HTML metadata processed by the paired agent during browsing sessions.
  • Boundary markers: The skill relies on technical capability restrictions (scopes) to limit impact rather than natural language prompt delimiters.
  • Capability inventory: Navigating to URLs, interacting with page elements, filling forms, taking screenshots, and executing JavaScript.
  • Sanitization: External content can be sandboxed by restricting JavaScript, cookies, and local storage access through CLI parameters.
  • [COMMAND_EXECUTION]: The skill executes various local binaries and shell commands within the ~/.claude/skills/gstack/ directory to manage browser state and tunnels. It includes specific security logic to prevent ngrok authentication tokens from being exposed in chat transcripts or shell history by instructing the user to perform authentication in their own terminal.
Audit Metadata
Risk Level
SAFE
Analyzed
Nov 16, 2026, 08:17 PM