Nebusec showed that a single malicious webpage could take over Meta Muse, the AI agent launched on September 8, 2026 (The Hacker News). In the lab test the team pulled a user’s private conversation history and stored memories without any security prompt (NebuSec).
Meta Muse runs inside an isolated cloud virtual machine that handles tasks, manages linked accounts, and makes purchases for users. Meta protects the agent with Sentinel, a monitoring layer that gates all outbound traffic and blocks direct access to sensitive account tokens (Muse).
The researchers linked a flaw in Muse’s built‑in browser to a Linux kernel vulnerability, giving them full control of the user’s isolated cloud instance.
With that control they accessed the credentials store, saved files and the Sentinel layer itself, and were able to alter stored memories and chat logs.
The breach stayed within the single user’s virtual machine; it did not affect the shared server or other users. Nebusec confirmed that all discovered weaknesses were reported to the Muse team and patched before public disclosure.
Meta has not commented publicly on the research nor confirmed the fixes. It remains unclear whether any malicious actors accessed user data before the patches were applied.
The successful exploitation of Meta Muse highlights the structural limits of relying on software-level sandboxing for autonomous AI agents. Even when architectures assume an agent will be targeted by prompt injection or malicious input, underlying infrastructure vulnerabilities like kernel flaws can render higher-level guardrails ineffective. As agents gain broader access to user credentials and system shells, isolating the execution environment from sensitive data stores will likely require hardware-enforced boundaries rather than reliance on logical monitoring layers like Sentinel.
Spotted an error? See our corrections policy or email editorial@lordofclaude.com.





Discussion & Comments (0)
Join the conversation