NebuSec Researchers Demonstrate Attack Chain Exposing Meta Muse Private Data

NebuSec Researchers Demonstrate Attack Chain Exposing Meta Muse Private Data

Cybersecurity researchers have demonstrated how a malicious webpage could exploit security weaknesses in Meta’s personal AI agent, Muse, to access private conversations and stored memories without triggering a user approval prompt. The research, published by cybersecurity firm NebuSec on November 20, 2026, showed that the attack could also be extended to reach credentials stored for users’ connected accounts. The findings raise concerns about the security boundaries protecting AI agents that interact with personal information and external online services. However, the demonstration was conducted in a controlled laboratory environment and did not establish that real users had been targeted or that their information had been stolen. NebuSec said it reported the weaknesses to Muse’s team and that all the exploits described in its research had been fixed before publication. Meta has not publicly confirmed the fixes or issued a separate comment on the findings. Muse, introduced by Meta on September 8, is designed to operate as a personal AI agent that can work across connected online accounts and perform tasks on behalf of users. Each user’s agent runs on a dedicated cloud-based virtual machine, with security controls intended to keep its activities and sensitive information isolated.

Meta designed Muse with several protections intended to limit the consequences of an agent being compromised. One of these is Sentinel, a monitoring layer responsible for controlling the agent’s outbound network communications and requesting user approval when required. The system is also designed to keep passwords and authentication tokens for connected services outside the agent’s direct access. However, NebuSec’s demonstration showed that these safeguards could be bypassed through a sequence of weaknesses. According to the researchers, a malicious webpage opened by Muse could exploit a vulnerability in its built-in browser, which is used for tasks such as generating PDFs and handling web content. This provided an initial opportunity to execute code within the agent’s runtime environment. The researchers then demonstrated that weaknesses in Sentinel’s network enforcement could allow conversation history and stored memories to be transmitted to an external server without triggering an approval prompt. They also showed that the information could be modified, creating risks to both the confidentiality and integrity of users’ stored data. The findings are significant because the exposed information was supposed to remain protected by separate security controls rather than being directly accessible to content loaded through the agent’s browser.

The research went further by demonstrating how the initial browser compromise could be combined with a Linux kernel vulnerability to gain control of the individual user’s cloud machine. NebuSec said it developed an exploit chain that allowed researchers to escape the agent’s runtime container and obtain root privileges within the user’s virtual machine. This level of access exposed resources that Muse’s security architecture was intended to keep separated from the agent, including saved files, the credential store and Sentinel itself. Access to credential storage could potentially expose passwords or authentication tokens associated with services connected to Muse, creating further risks beyond the initial disclosure of conversations and memories. Taking control of Sentinel also meant that the researchers could reach resources and network capabilities that were protected by its monitoring and enforcement mechanisms. However, the researchers made an important distinction about the scope of the demonstration: the compromise remained inside the individual user’s virtual machine. They did not demonstrate an escape into the underlying shared server or access to other users’ machines. This limitation is important when assessing the findings, as the research showed a serious compromise of an individual environment rather than a breach affecting the entire Muse infrastructure or its user base.

NebuSec’s findings highlight the challenges of securing AI agents that require access to personal data, connected accounts and online services to complete tasks. Meta’s security approach was designed around the possibility that an agent could be manipulated or compromised, using isolation, credential separation and network monitoring to limit potential damage. The research nevertheless demonstrated how weaknesses across different components could be chained to undermine several of these protections. The issue also illustrates why securing an AI model or restricting its direct permissions may not be sufficient when the surrounding browser, operating system, network controls and credential management systems remain part of the same execution environment. NebuSec said it disclosed the findings to Muse’s team and that the vulnerabilities used in the demonstration had been addressed before its report became public. The researchers did not establish whether any real user information had been accessed before the fixes were implemented. The disclosure follows an earlier Muse security issue reported in September, which involved malware already present on a user’s Mac hijacking the assistant locally. Unlike that earlier issue, the latest research examined the cloud infrastructure behind Muse and demonstrated a route from malicious web content to sensitive resources within an individual user’s virtual machine. For organizations developing or deploying AI agents, the findings underline the importance of testing security boundaries across the entire platform, including the components responsible for isolating agents and protecting user credentials.

Source

Follow the SPIN IDG WhatsApp Channel for updates across the Smart Pakistan Insights Network covering all of Pakistan’s technology ecosystem.

Post Comment