What happened
Security firm NebuSec demonstrated that a booby-trapped webpage opened by Meta Muse, the company's personal AI agent, could copy a user's conversation history and saved memories to an outside server and change them, with no approval prompt ever appearing, The Hacker News reported. The work was a laboratory exercise and did not target real users.
The researchers then combined a weakness in the agent's built-in browser, which Muse relies on for jobs such as producing PDFs, with a bug in the Linux kernel to gain complete control of the individual user's cloud machine. That control extended to the credential store for connected accounts, saved files and Sentinel, the monitoring layer meant to oversee all outbound agent activity. The attack did not reach the shared server underneath or other users' machines.
The firm said it had shared its findings with the Muse team and that all of the exploits were fixed ahead of publication. Meta has issued no public confirmation of the patches or any comment on the findings, and the write-up leaves open whether user data was exposed before the fixes landed.
Context
Meta launched Muse on September 8 as an agent that operates through linked accounts and can buy things for users, each on its own cloud computer. Its design assumed the agent itself might be compromised and aimed to limit the fallout, keeping account passwords and tokens out of the agent's reach. An earlier September flaw was narrower, letting malware already on a user's Mac take over the assistant locally.
ब्रीफ़िंग
टेक का पूरा सप्ताह, एक ईमेल में।
हर सप्ताह वे रिपोर्ट्स जो आपके समय के लायक हैं। कभी भी अनसब्सक्राइब करें।
