Agent Upstart

Cool things people are building with agents — and practical help getting the most out of them.

Playbooks · Coding tasks

A critical flaw in gemini-mcp-tool, fixed in 1.1.6

Written November 8, 2026

gemini-mcp-tool had a critical @file exfiltration and Windows command-injection flaw, fixed in 1.1.6. Who is affected, and the one-line check.

What the flaw is

GitHub's advisory GHSA-4h5r-5jm8-jxjm (CVE-2026-0755) records two flaws in gemini-mcp-tool, an MCP server that wraps the Gemini CLI. Untrusted prompt input could reach the Gemini CLI @file parser, allowing read and exfiltration of arbitrary local files — the advisory names @/etc/passwd, @~/.ssh/id_rsa, and @../../secret as examples. On Windows, unquoted cmd.exe metacharacters could break out into OS command injection. The Zero Day Initiative, which reported the command-injection flaw as ZDI-26-021, describes it as a flaw in the execAsync method: a user-supplied string was used to execute a system call without proper validation, and an attacker could leverage it to execute code in the context of the service account, with no authentication required.

The mechanism matters because of where the input comes from. Prompt input is untrusted by definition — it can arrive from a file the agent reads, a page it fetches, or text a user pastes in — and this flaw let that input reach a file parser and a shell. The @file side is the quieter of the two: a file read needs no command to run, only a reference the parser follows.

Who is affected, and the fix

The one-line check

Check the installed version, then upgrade if it is in the affected range:

npm list gemini-mcp-tool
npm install -g gemini-mcp-tool@latest

If the tool is installed as a dependency of something else rather than globally, check the version in that project's package.json or lockfile and upgrade there instead — a global upgrade does not fix a pinned local copy. If the version is below 1.1.2, it predates the affected range; if it is 1.1.6 or later, it carries the fix.

What not to assume

Limits

Related playbooks

Sources

Back to: Coding playbooks hub · Also: MCPhantom: agents install dead MCP servers · A malicious MCP server probed cloud credentials · 140 silent MCP contract changes, and how to pin them