OpenCode on Windows or WSL: can a remote MCP server run commands during 'opencode mcp auth' sign-in? Which version fixes it?

Yes, on affected versions. When OpenCode signed in to a remote MCP server with OAuth, it opened the authorization URL supplied by that server through a helper that didn't escape it properly on Windows and WSL, so a malicious MCP server could get commands run with your user's privileges during sign-in. No CVE was assigned. Am I affected: opencode-ai 1.0.137 through 1.18.32, running on Windows or WSL, and you've signed in (or will sign in) to a remote MCP server you don't fully trust. macOS and Linux aren't affected. Fix: 1. Upgrade OpenCode to 1.18.33 or later (released 2026-09-28). 2. Review configured servers with opencode mcp list; remove any remote MCP server you don't recognize or trust, and sign out of it with opencode mcp logout <name>. 3. Where you don't need OAuth for a server, disable it in its config ("oauth": false). If you signed in to an untrusted MCP server on an affected version, treat that machine as possibly compromised: check for unexpected processes and scheduled tasks, and rotate credentials stored on it (API keys, git and cloud tokens). Sources: https://thehackernews.com/2026/10/opencode-flaw-lets-malicious-mcp.html ; https://safeguard.sh/resources/blog/opencode-console-key-breach-and-mcp-oauth-rce