Security firm NebuSec showed in a lab that a webpage opened by Meta's Muse agent could pull out and alter a user's chat history and stored memories without any approval prompt, and by chaining a browser and a Linux kernel flaw, take over the agent's cloud machine including its credential store. NebuSec says Meta fixed the issues before publication; Meta hasn't commented.
Why it matters: Before you let an AI agent browse the web with access to company accounts, give it its own low-privilege accounts and keep payment and admin credentials out of its reach; an approval layer alone isn't enough.