"So we designed the system to assume the agent may be under attack and limit the potential damage," Muse's team wrote at launch — a design principle that a new laboratory demonstration nevertheless managed to bypass.
NebuSec's laboratory demonstration and the exploit chain
Security firm NebuSec published a technical write-up showing that a single malicious webpage opened by a Meta Muse agent can be the starting point for a multi-stage compromise. In the lab test the researchers chained a flaw in Muse's built-in browser, which the assistant uses for tasks such as creating PDFs, to a Linux kernel flaw. That chain gave the researchers full control of the individual cloud machine that runs a user's personal Muse agent.
What the malicious page could extract and alter
NebuSec's report says the malicious page could extract a Meta Muse user's private conversations and stored memories from the assistant and send them to an external server without any approval prompt being shown to the user. From that foothold the researchers say they were also able to reach the credentials Muse stores for the user's connected accounts. The write-up adds that the page could not only pull data out but also alter conversations and stored memories inside the agent.

Audit-ready is a season. It shouldn't be.
Evidence in spreadsheets, controls drifting between audits, frameworks multiplying on flat headcount. Nubivance runs continuous compliance on Rapid7 Cyber GRC - SOC 2, HIPAA, ISO 27001, PCI, CMMC.
End the scrambleHow Sentinel and other defenses fared
Muse was architected so each person's agent runs on its own cloud computer and is monitored by a layer called Sentinel, which is intended to be the single gate for anything the agent sends out. Passwords and tokens for connected accounts are supposed to remain beyond the agent's reach. According to NebuSec, those controls "worked as designed" in terms of containment, yet the researchers were still able to carry the attack through to full control of the user machine and into the parts of the system the design was meant to keep walled off.
Scope of the compromise: limited to one cloud machine
The researchers emphasize that the demonstration stayed inside the single user's cloud machine. NebuSec reports that it did not break out to the shared server beneath that machine or into other users' machines. The firm notes this compromise therefore did not produce lateral movement across the Muse infrastructure. The write-up also contrasts this lab work with an earlier Muse flaw discovered in September, which only allowed malware already on a user's Mac to hijack the assistant on that device.
What this means for technologists, policymakers, and end users
- Technologists and security teams: will likely focus on browser-to-kernel attack surfaces and the interactions between Muse's built-in browser, the local guest environment, and kernel-level vulnerabilities. They will also need to verify that the credential store and Sentinel can now withstand chained exploit paths that start with web content opened by an agent.
- Policymakers and regulators: should note that personal AI agents run on per-user cloud machines and can access account connections and stored memories — vectors that, if abused, allow extraction of sensitive information without in-product prompts. NebuSec's finding that such an extraction was possible without user approval will be a concrete example for oversight conversations.
- End users and privacy-conscious buyers: will want clarity on whether any user data was accessed in the wild. NebuSec says it reported the findings to Muse's team and that the exploits were fixed before the firm's publication, but Meta has not publicly confirmed the fixes or commented, and the write-up does not say whether any user data was accessed before patches were applied.
The demonstration underscores a practical tension in the design of personal AI agents: they must be powerful enough to act on a user's behalf, including accessing connected accounts and stored memories, yet constrained enough to prevent those same capabilities from being weaponized by content the agent itself loads. NebuSec reports that every weakness it described was fixed after responsible disclosure; Meta has not publicly confirmed those fixes or offered comment, and the write-up leaves open whether any real user data was exposed prior to remediation.
For now, the concrete facts from NebuSec's write-up are straightforward: a malicious webpage opened by a Muse agent can extract and alter private conversations and memories without prompting, and when combined with a browser flaw and a Linux kernel flaw in sequence it can give an attacker full control of the user's cloud machine — reaching credential stores, saved files, and Sentinel itself — yet stopping short of the shared infrastructure. Whether that containment model can be relied on in the long run, and how quickly independent verification of the fixes will appear, are the next questions that remain.




