Claude AI Found a Critical Vim Zero-Day in Under Two Minutes

A security researcher handed Anthropic’s Claude Code a single instruction: “Somebody told me there is an RCE 0-day when you open a file. Find it.”

Two minutes later, Claude had identified a critical remote code execution vulnerability in Vim — the text editor installed on virtually every Linux server, Mac, and developer workstation on the planet. The bug, now tracked as CVE-2026-34714, allows an attacker to execute arbitrary operating system commands simply by tricking someone into opening a crafted file.

No plugins. No macros. No confirmation dialog. Just open the file.

The Vulnerability: A Two-Link Chain

The exploit chains together two separate oversights in Vim’s codebase that, combined, break out of the editor’s security sandbox:

Link 1: Missing Security Flag

Vim’s tabpanel option — part of the sidebar feature introduced in version 9.1.1391 — was missing the P_MLE (Prevent Modeline Expressions) security flag. Similar options like statusline and tabline have this flag. The tabpanel option simply never got it.

This means a malicious file can set tabpanel to contain a %{expr} expression that Vim will evaluate when the file is opened.

Link 2: Sandbox Escape

Vim correctly evaluates these expressions inside a sandbox. But the autocmd_add() function was missing a check_secure() call. This means sandboxed code can register an autocommand that fires after the sandbox exits, with full privileges. Game over.

The Attack Flow

1. Attacker crafts a file with a malicious modeline
2. Victim opens the file in Vim — modelines are enabled by default
3. The modeline injects a %{expr} expression into the tabpanel setting
4. The expression runs inside the sandbox and calls autocmd_add()
5. The registered autocommand fires after the sandbox exits
6. Arbitrary OS commands execute with the victim’s privileges

The Numbers

Who Found It

Discovered by Hung Nguyen from Calif, an AI red teaming company, using Claude Code. Part of the MAD Bugs initiative (April 2026). Claude identified exact source locations:

• src/optiondefs.h:2581 — Missing P_MLE flag
• src/option.c:1572-1576 — Modeline security checks bypassed
• src/autocmd.c:3316 — autocmd_add() lacks check_secure()

Project Glasswing

Anthropic announced Project Glasswing, a $100M initiative with 12 partners (AWS, Apple, Microsoft, Google, NVIDIA, Linux Foundation) to find vulnerabilities in critical software.

• Thousands of high-severity zero-days across every major OS and browser
• A 27-year-old flaw in OpenBSD
• A 16-year-old FFmpeg vulnerability undetected across 5M test runs
• 72.4% exploit success rate for Mythos Preview

“We did not explicitly train Mythos Preview to have these capabilities. Rather, they emerged as a downstream consequence of general improvements in code, reasoning, and autonomy.”

What You Should Do

Update Vim to 9.2.0272+. Temporary fix: add set nomodeline to your .vimrc.

The Uncomfortable Question

An AI found a critical zero-day in one of the most widely-deployed editors on Earth in under two minutes. It had been hiding in a codebase audited by humans for over three decades. If Claude can find this, what else is out there?

Sources: GitHub Advisory | Anthropic Glasswing | Help Net Security | The Register