John Z Black
Brunswick, ME • (207) 245-1010 • contact@johnzblack.com
A security researcher pointed Claude at the Vim and Emacs codebases. It found file-open RCE vulnerabilities in both. Open a file, get owned. Classic.
Vim’s maintainers patched it immediately. Version 9.2.0272 fixes the modeline sandbox bypass. Responsible, fast, done.
Emacs? The maintainers said it’s Git’s problem. It isn’t.
The Emacs bug works through vc-git integration. Open any file in a directory with a crafted .git/config, and Emacs will execute whatever command is in the core.fsmonitor setting. No prompts, no warnings. Stick a malicious .git folder inside a ZIP file, send it to someone, and when they open any file from that directory in Emacs? Game over. No CVE assigned. No patch coming.
Saying “Git should fix it” when your editor is the thing executing arbitrary commands from an untrusted config is a bold move.
Then there’s the bigger picture. Check Point published details on an unreleased Anthropic model codenamed “Mythos” that reportedly shows a significant jump in offensive security capabilities. Multi-step attack chain reasoning approaching what you’d expect from nation-state teams. The gap between a well-resourced APT and a solo researcher with model access is shrinking fast.
These two stories together tell you where things are headed. AI can now find bugs in codebases that humans have audited for decades. And some maintainers still won’t patch what it finds.