News Security news
Meta Muse: a single malicious webpage could pull chats and memories out of the AI assistant
On 20 November 2026, researchers at NebuSec published how a single malicious webpage let them take over the environment of Meta Muse, a cloud-hosted AI assistant. The attack sent a user's conversation history and stored memories to an external server without any approval prompt and, combined with a Linux kernel flaw, gave root on the user's virtual machine. According to NebuSec, the Muse team fixed every issue before publication.
Key points
- The entry point was a flaw in the Chromium browser Muse uses for tasks such as generating PDFs and documents, which gave code execution inside the agent's container.
- Weaknesses in Sentinel, the service meant to police the agent's outbound traffic and ask for approval, let sensitive data out without a prompt.
- A Linux kernel flaw then allowed a container escape to root on the user's VM, exposing persistent storage, the credential store and Sentinel itself.
- The compromise was limited to one VM. NebuSec gives no CVE numbers or report and fix dates; Meta has not commented publicly.
Security firm NebuSec has published research in which it extracted conversations and memories from Meta Muse, a cloud-hosted AI assistant, and eventually gained full control of the user's virtual machine. All it took was for the agent to open a malicious webpage. The researchers say every flaw has been fixed.
What happened
Muse runs in an isolated cloud VM, with the agent further separated by a runtime container. The researchers chained three weaknesses:
- The browser. A flaw in the Chromium instance Muse uses for web tasks and document generation gave them code execution inside the agent's container.
- The outbound traffic guard. Sentinel is meant to decide which outbound requests need the user's approval. Flaws in how it classified and inspected requests let conversations and memories go to an external server with no prompt at all. The attacker could also modify them.
- The Linux kernel. A flaw in the shared kernel (NebuSec does not name it) let them escape the container to root on the VM - reaching persistent storage, the store holding credentials for connected accounts, and Sentinel itself.
Who is affected
Meta Muse users - but the fixes were made on Meta's side and this was a lab demonstration, not an attack on real users. The research does not say whether anyone abused the flaws earlier; no exploitation has been reported.
Why it matters
The research shows a pattern we keep seeing with AI agents: the agent opens untrusted content, and the safety net - approval for outbound actions - rests on a single service. When that fails, data leaves without the user knowing. We covered a related trend with Cairn, where AI agents carried out attacks on their own.
What to do
- Muse users: based on what has been published, there is nothing you need to do. If you want to be thorough, review the assistant's stored memories and disconnect accounts you no longer need.
- Teams deploying AI agents: treat every page and document an agent opens as untrusted input; isolate and patch the agent's browser separately; don't let approval for outbound actions depend on one control; give the agent only the credentials it really needs, with short lifetimes.
- Check suspicious links with our Link checker before handing them to an AI assistant.
Sources
Spotted a mistake? Let us know and we will correct it openly.