News
A malicious webpage could read a Meta Muse user's chats without a prompt.
Researchers showed that a malicious page opened by Meta Muse could copy a user's conversation history and stored memories, and change them, with no approval prompt. The same chain reached credentials stored for connected accounts. It was a lab test on one user's cloud machine, and the firm that published it says Muse's team has fixed the weaknesses. Meta has not confirmed that.

Impact
- What was impacted
- Meta Muse, the personal AI agent Meta launched on September 8. Each person's agent runs on its own cloud computer. A monitoring layer called Sentinel is meant to be the only gate for anything the agent sends out, and passwords and tokens for connected accounts are supposed to stay out of the agent's reach. In the lab, a malicious page the agent opened pulled conversation history and stored memories to an outside server and changed them, with no prompt. The researchers then chained a flaw in Muse's built-in browser, the one used for tasks such as creating PDFs, to a Linux kernel flaw and took control of that user's cloud machine, including the credential store, saved files, and Sentinel.
- Who was impacted
- People who use Meta Muse, and the accounts they have connected to it. The demonstration stayed inside one user's cloud machine. The researchers did not break out to the shared server under it or into other users' machines. NebuSec, which published the work yesterday, says this was a laboratory demonstration, not an attack on real users. The write-up does not say whether any user data was read before the claimed fixes. Meta has not publicly confirmed the fixes or commented. An earlier Muse flaw, found in September, was a different case: it only let malware already on a person's Mac hijack the assistant on that device.
- What you need to do
- This report names no Muse version, no CVE, and no setting to change. NebuSec says it reported the findings and that the exploits were fixed before publication. Meta has not confirmed that. Do not treat the issue as closed on the firm's statement alone, and do not invent an install step this report does not give. Because the lab chain reached stored credentials, know which accounts are connected to Muse. This report does not say any real user's data was taken, so do not rotate those credentials unless you already have a separate reason to. Do not try to reproduce the page, the browser flaw, or the kernel chain.
A malicious webpage could pull a Meta Muse user's private conversations and stored memories without an approval prompt. The same chain could reach credentials Muse keeps for accounts the user has connected. NebuSec published the work yesterday. It was a laboratory demonstration, not an attack on real users, and it stayed inside that one user's cloud machine. The firm says Muse's team has fixed every weakness the write-up describes. Meta has not confirmed the fixes.
Muse works through online accounts a person connects, and it can make purchases on their behalf. Each agent runs on its own cloud computer. Sentinel is supposed to be the single gate for anything the agent sends outward, and the passwords and tokens for connected accounts are supposed to stay beyond the agent's reach. At launch, Muse's team said the system was built on the assumption that the agent might be attacked, and that the design should limit the damage.
In the demonstration, a malicious page the agent opened copied conversation history and stored memories to an outside server and changed them. No prompt was shown. From there the researchers chained a flaw in Muse's built-in browser, which the assistant uses for work such as creating PDFs, to a Linux kernel flaw, and took control of the user's cloud machine. That control included the credential store, saved files, and Sentinel. The break did not leave that machine. The researchers did not reach the shared server underneath it or other users' machines.
The firm says those controls worked as designed, and the attack still ran the whole way through. It says it reported the findings and that the exploits were fixed before publication. The write-up does not say whether any user data was read before those claimed fixes, and Meta has not commented. This is a break into the cloud system behind the assistant. The Muse flaw reported in September was narrower: malware already on a Mac could hijack the assistant on that device.
There is no version number and no menu in this report. The only fix status on the page is the research firm's statement. Do not rebuild the page or the chain to see whether a machine is still exposed.
For your AI agent
Agent instructions — copy paste to Claude
You are helping me respond to a laboratory demonstration against Meta Muse. A malicious webpage opened by the agent could copy a user's conversation history and stored memories, and change them, with no approval prompt. A further chain took control of that one user's cloud machine, including stored credentials. Do not write or run a webpage, a browser exploit, a kernel exploit, or a proof of concept. Do not ask Muse to open a page in order to test this. Ask before disconnecting an account, rotating a credential, or changing a Muse setting. Facts from the report: - NebuSec published the work yesterday. It was a laboratory demonstration, not an attack on real users, and it stayed inside the individual user's cloud machine. Do not describe it as a breach of real customer accounts. - A malicious page the agent opened could pull conversation history and stored memories to an external server and alter them, with no prompt shown to the user. The same attack could be extended to credentials Muse stores for connected accounts. - Muse is Meta's personal AI agent, launched on September 8. It works through online accounts a user connects and can make purchases on their behalf. Each person's agent runs on its own cloud computer. - A monitoring layer called Sentinel is meant to be the single gate for anything the agent sends out. Passwords and tokens for connected accounts are supposed to stay beyond the agent's reach. Muse's team said at launch that the system assumes the agent may be under attack and should limit the damage. Paraphrase that. Do not present it as a new control I can turn on. - The researchers chained a flaw in Muse's built-in browser, used for tasks such as creating PDFs, to a Linux kernel flaw, and took full control of the user's cloud machine. That control reached the credential store, saved files, and Sentinel. Do not reconstruct those steps. - The compromise did not break out to the shared server beneath that machine or into other users' machines. The firm says each of the designed controls worked, and the attack still ran the whole way through. - The firm says it reported the findings and that the exploits were fixed before publication. Meta has not publicly confirmed the fixes or commented. The write-up does not say whether any user data was accessed before the patches. Keep both limits. Do not tell me the issue is closed. - This report names no CVE and no Muse version. Do not invent either. - An earlier Muse flaw found in September only let malware already on a user's Mac hijack the assistant on the device. Do not mix that case into this cloud-machine chain, and do not invent steps for it. Context: - What was impacted: one Muse user's cloud machine in a lab, including chats, stored memories, and the credential store for connected accounts, plus Sentinel. - Who should care: people who use Muse and have connected other accounts. The demo did not cross into other users' machines and was not described as an attack on real users. - What to do: there is no version, CVE, or setting in this report. The firm says the weaknesses were fixed before publication. Meta has not confirmed that. Do not rotate credentials unless I already have a separate reason to. Do this in order. Ask before changing an account or a setting. Prefer read-only inventory first. Do not write or run an exploit or a proof of concept. 1. Inventory - Record whether I use Meta Muse, and list the accounts I have connected to it if that list is visible without changing anything. If you cannot see the product, say so and stop. - Record whether I have a statement from Meta that these weaknesses are fixed. A statement from the research firm is not that confirmation. Do not invent a Meta advisory. - This report names no version. If a version is visible in the product, write it down and say this report cannot tell you whether that build is the fixed one. 2. Contain - Do not tell me to disconnect every account. The report does not say real user data was taken. - If I want to pause new connections until Meta confirms a fix, propose that and ask before changing anything. 3. Remediate - There is no patch version and no setting in this report. Do not propose an update, a menu, or a configuration change the report did not name. - Do not rotate a connected-account password or token unless I tell you I already have a separate reason. If I do, ask before changing it. 4. Verify - If I asked you to change a connection, read that connection list again. - If I did not ask you to change anything, report the inventory only. 5. Report back - Whether I use Muse, which connected accounts you could see, whether Meta has confirmed a fix in anything I showed you, and that the demonstration was a lab test of one cloud machine. Do not invent CVEs, Muse versions, victim counts, or signs that a particular chat was read. If a check needs access I have not given you, say what you need and stop.