The Hacker News
Malicious webpage exposed Meta Muse chats and memories
A lab demonstration showed that a single malicious webpage could access a Meta Muse user’s private conversations and stored memories from the assistant without triggering any approval prompt. The researchers demonstrated that, from within Muse’s built‑in browser used for tasks like generating PDFs, the attack could pull conversation history and memories to an external server and even alter them, all without user interaction.
The chain of compromise extended to the Linux kernel through a flaw linked to Muse’s cloud environment, enabling full control of the affected user’s cloud machine. NebuSec, the security firm behind the disclosure, notes Muse’s team has since fixed every weakness described, and that the attack did not escape the initial machine or reach other users’ devices.
The project was explicitly described as a laboratory demonstration and not an attack on real users; Meta has not publicly confirmed the fixes or commented on the research, and the write‑up does not state whether any user data was accessed before patches were applied.
The researchers provide a technical breakdown of how the flaw chain operated and state that the compromises remained contained to the single host cloud machine. NebuSec’s write‑up indicates that Muse’s security controls—designed to wall off credentials, saved files, and Sentinel—bound the attack, even though the attack traversed the chain to that machine’s credential store and memory.
The disclosure follows an earlier Muse flaw reported in September 2026; there is no indication in the material that any real users were harmed, and Meta has not issued public confirmation of fixes. In the public write‑up, the Linux kernel flaw cited is CVE‑2026‑99312.