Laboratory test: one malicious webpage was enough to extract a Muse user’s chats and memories
Security company NebuSec says that in a laboratory test it sent a Muse user’s chats and memories to an outside server without an approval prompt, and eventually took control of the user’s cloud machine. Meta fixed the flaws before publication, the company said.
Muse is a personal AI agent that Meta launched on 8 September. It works through the online accounts a user connects to it and can make purchases on the user’s behalf. NebuSec described the attack in a research post published on 20 November.
Muse has layered protection. Each user’s agent runs in its own virtual machine in the cloud, inside a separate container. A service called Sentinel monitors outbound traffic and asks the user for approval when needed. The credentials for connected accounts are meant to stay out of the agent’s reach.
The researchers did not trick the language model in conversation. Instead, the attack started from one of the agent’s own tools. Muse uses the Chromium browser for tasks such as generating PDFs. When the agent converted the attacker’s page into a PDF, the researchers exploited a browser vulnerability to run their own code in the container.
NebuSec said Sentinel did not properly classify and check outbound requests. This let the researchers send the chat history and memories to an outside server without an approval prompt. They could also alter the agent’s memories and chat history.
Next, the researchers exploited a Linux kernel vulnerability to escape the container and gain root privileges on the virtual machine. NebuSec’s automatically generated exploits succeeded in more than 99.9 per cent of test runs. With root privileges, the researchers gained access to storage, the secure credential store and Sentinel. FourWeekMBA notes that NebuSec does not say it extracted any credentials. The compromise stayed within one user’s virtual machine, and the researchers did not demonstrate access to the server running the virtual machine or to other users’ machines.
NebuSec said the protections worked as designed, but each step in the chain weakened a different layer, so the attack went all the way through. The company urges agent developers to check whether their protections still hold once a system is partly compromised. Credential permissions should also be restricted in case of a breach, it said.
This was a laboratory test, not an attack on real users. NebuSec says it reported its findings to Meta and that Meta fixed all the flaws described before publication. The Hacker News reported that Meta has not publicly confirmed the fixes. The post does not make clear whether anyone accessed real users’ data before the fixes. The case is separate from a Reuters report, summarised by TechCrunch, that Meta employees had raised concerns in internal messages that Muse was mixing up conversations from different accounts.
UrduPoint advises Muse users to update the app and check their connected accounts. It is also worth reviewing how freely the agent may browse the web and make purchases, and removing unnecessary connections. NebuSec has not said whether the findings also apply to Muse’s Chrome extension, FourWeekMBA reported.