A malicious webpage was all it took in a NebuSec demonstration to access Meta Muse's conversation history and saved memories, without showing the user an approval prompt. The researchers chained flaws to take control of an account's cloud virtual machine; NebuSec says Meta fixed the issues before the research was published.
The research affects people who use Meta Muse, Meta's AI agent, which operates with connected accounts and can perform tasks on the user's behalf. This does not mean all those accounts were compromised: NebuSec says the intrusion remained within the individual test virtual machine and did not spread to other users or the shared server. The company reported no real-world attacks and could not determine whether user data had been accessed before the fixes were applied.
What was demonstrated and what you can check
The chain began when Muse opened a page controlled by an attacker. According to NebuSec, a flaw in the integrated browser allowed code to run inside the agent's container; then, weaknesses in Sentinel's traffic inspection made it possible to extract conversations and memories without triggering the warning. A Linux kernel flaw shared between the container and the virtual machine opened access to storage, Sentinel, and the credential vault.
The research was a lab test, not an alert that a specific page is currently attacking users. NebuSec says it reported the findings to the Muse team and that the exploits described had been fixed before publication. Meta had not publicly confirmed those fixes or commented on the research; it is also unknown whether data was accessed before the patches were applied.
The reports describe components of Muse's cloud infrastructure, but do not provide a procedure for checking whether a specific account was exposed. They also do not identify a Muse version that would let users verify the fix on an individual account.
Steps to reduce exposure
Don't use Muse to open suspicious pages until you have clear confirmation that the fix has been deployed for your account. The demonstration required the agent to process web content controlled by an attacker. Don't open unexpected links in Muse or ask it to export or convert pages you don't consider trustworthy. NebuSec says the flaws were fixed, so this is a precaution, not a claim that the attack still works.
Avoid entrusting it with especially sensitive data until your questions about the service are resolved: don't ask it to process private conversations, sensitive documents, or login details. The test managed to extract stored history and memories; after escalating to control of the virtual machine, it also reached the credential store. This describes what the researchers achieved in their environment, not confirmation that this data was leaked from real accounts.
Review which services you have connected and what access Muse actually needs for the tasks you assign it. The research describes credentials and tokens for connected accounts, but does not provide a Muse menu path for revoking permissions. If you decide to remove access, do so through the official controls of the connected service or Meta that are available to you; don't follow third-party instructions promising a specific option you can't verify.
Check Meta's notices about the fixes before delegating sensitive tasks again. NebuSec reported that the weaknesses had been fixed by the time of publication, while Meta had not publicly confirmed this in the information available. If the service offers support or security information, ask specifically whether the fix covers your account's instance and whether any user action is recommended; the researchers did not describe an individual verification procedure.
Take action if you see activity you don't recognize on connected accounts. Check those services directly and change the relevant credentials if you find unfamiliar logins or actions; don't reuse passwords. The research does not report any real credential theft, but it did demonstrate that the chain could reach the virtual machine's credential store. Keep any notices or logs you find and contact the affected service through its official channels.
Distinguish this finding from other incidents attributed to Muse. ZDNET cites a Reuters investigation into unexpected behavior during employee tests and tests of access to messages through existing permissions on Android and Mac; the cited tests did not work on iPhone. These are separate issues from NebuSec's chain, which began with a malicious page opened by Muse and compromised its cloud environment. Don't conclude that a phone permission proves your virtual machine was attacked, or vice versa.

Mistakes to avoid
Don't interpret the demonstration as a confirmed intrusion into your account. NebuSec explicitly says its test did not leave the individual virtual machine and did not reach other users. There is also no evidence to support claims that anyone exploited these flaws against real users.
Don't assume the approval prompt, Sentinel, or the agent's isolation were enough to prevent the chain: the researchers describe how they bypassed those safeguards. But don't conclude that all of Muse's security systems remain vulnerable, either; NebuSec says the flaws described were fixed before the research was made public.
Finally, don't look for a specific Muse version to certify that you're safe: the reports do not identify a Muse version that would let users verify the fix on their account. The specific question that remains open is whether Meta can confirm that the fixes have been deployed to every account and whether it detected any access before the patches were applied.
Sources: The Hacker News, ZDNET, la Repubblica, NebuSec, NebuSec