One webpage turned Meta's AI agent against its own user. What logins have you handed your AI assistant? I want to be clear up front: this was a laboratory demonstration, not a break-in that hit real customers. Researchers at NebuSec showed that a single malicious webpage was enough, in their test, t
o read a Muse user's private chats and stored memories, with no approval prompt shown to the user. They then reached the saved logins for the accounts that user had connected. The work stayed on that one test user's cloud machine. It did not reach other users, and it did not reach the shared servers underneath.
Muse is Meta's personal AI agent. The Hacker News, writing on 21 November 2026, says it launched on 8 September, that it works through the online accounts a user connects to it, and that it can make purchases on their behalf. Each person's agent runs on its own cloud computer. NebuSec published its write-up on 20 November 2026.
In the demonstration, the agent opened a malicious page. From there, the researchers could pull the user's conversation history and stored memories to an outside server, and change them, without any prompt being shown to the user. They then took full control of that one user's cloud machine. That included the store of saved logins for connected accounts, and the files saved there.
I am not going to walk through how they did it. The useful point for a business owner is what the agent was already trusted to hold. NebuSec says it reported the findings to Muse's team, and that everything it describes had been fixed by the time it published. Meta has not publicly confirmed those fixes, and it has not commented on the research.
The write-up also does not say whether any real user data was accessed before the patches went in. I am leaving that as an open question, not filling the gap. This was a test on one user's machine. It is not a notice that Muse customers were breached, and I have not seen a public statement from Meta in the sources I opened.
The people who should still pay attention are owners who have already handed an AI assistant their logins: hosting, store admin, email, payments, or social accounts. If you connected those accounts so the assistant can "just handle it," this is about that habit, even if you have never used Muse. You give an assistant a login because it is convenient. On a normal day that feels like setup.
In this test, a page the agent itself opened was enough to turn that access against the user, with no approval prompt. The private chats and memories matter. So do the saved logins. If the assistant can reach your store admin, your email, or a payment account, a page it visits can become a problem for the business, not only for the chat history.
I have written about this kind of trust before. What to check after research on AI agent permissions. An AI assistant connector opened 56 projects. Hackers' AI found a hole nobody knew about and broke in within 11 minutes.
Different incidents. Same question: what did you already hand over? If it does not need your store admin, your bank, or your email, do not connect them. Go through the accounts already tied to each assistant you use, and remove the ones you added "just in case." A tool you tried last month can still have access today.
If the tool can work with a narrower key, use that. A full admin login is a full set of keys. If you cannot get a narrower option, ask whether you still need that assistant connected at all. If it can browse, send, buy, or change something with no prompt, write that list down. Cut anything you are not comfortable leaving on autopilot. In this test, there was no approval prompt before private chats and memories were taken. If you would not open a stranger's link while signed in to your store or hosting panel, do not let the assistant do that either. Finish the admin work, or disconnect those accounts, before you let it wander.
After you review those accounts, scan your website to see what an outsider can reach from the internet, without logging in. A free website security scanner can show what an outsider can see on your own website, without logging in. It will not audit the AI assistants you have connected, and it will not by itself prove you are safe. It is a starting point for the part of the risk that sits on your site.
If a website security scan finds a serious issue, our guide to what to do when a website scan finds critical flaws is a place to start. A few important things are still open, and I would rather say so than fill the gaps. Meta has not publicly confirmed the fixes, and it has not commented on the research in the sources I opened. The write-up does not say whether any real user data was accessed before the patches went in.
If any of that changes, the useful steps stay the same: connect only the accounts the assistant needs, prefer limited keys, review what it can do without asking, and do not let it browse untrusted sites while it is logged in to your admin accounts. No. Both sources describe a laboratory demonstration. The work stayed on one test user's cloud machine.
It did not reach other users, and it did not reach the shared servers underneath. I am not treating this as a breach of real customers. They could read the user's private chats and stored memories, and change them, with no approval prompt. They then reached the saved logins for connected accounts, and the files saved on that one user's cloud machine.
NebuSec says it reported the findings to Muse's team, and that the weaknesses it describes had been fixed before it published. Meta has not publicly confirmed those fixes or commented on the work, in the sources I opened. That is the question I would start with this week, even if you do not use Muse. Make a simple list of every AI assistant or agent that can act in your accounts.
For each one, ask which logins it has, whether it still needs them, whether it can act without asking you, and whether it can browse the web while those logins are connected. If you cannot answer those points, treat that as a problem until you can.
Key Takeaways #
- •They then reached the saved logins for the accounts that user had connected.
- •This story was reported by Dev.to , covering developments in thedev space.
- •AI advancements continue to reshape industries — read the full article on Dev.to for complete coverage.
📖 Continue reading the full article: