Meta’s Muse AI agent, launched in September, has racked up more than 10 million downloads — but even as it becomes one of the fastest-growing consumer AI tools, serious privacy concerns have emerged from both users and internal staff.
Muse is now under scrutiny for leaking private and unrelated data, including messages and financial info allegedly belonging to other people. Multiple employees have reported that Muse pulled in sensitive bank account balances they never shared, surfaced conversations from apps they hadn’t connected, and even interfered with messages in apps like Signal — all without explicit permission. Reuters and internal feedback forums show these issues go back to mid-September and continue today. Meta describes them as “isolated incidents,” attributing many to AI hallucinations. Staff are concerned those errors may stem from flawed design and permission settings.
The Zero-Day and Cross-Account Fears
Alongside these leaks, a zero-day vulnerability was discovered affecting Muse on Mac. This flaw reportedly allowed attackers to hijack voice requests and intercept authentication tokens, putting user data at risk. Meta has reportedly issued a safety warning in response.
Staff also describe episodes where Muse accidentally merged contexts between different users. For example, Muse offered career advice never requested, suggested purchases never made, and added bank info in unrelated chats. One employee said it reminded them to reply to a Signal message from their sister despite not giving Muse access to that app. Another said Muse added content about buying a coffee table that didn’t belong to them — an apparent mix of cross-account data.
Former Security Head Sounds the Alarm
Alex Stamos, Meta’s former chief security officer, has publicly warned about the risks Muse poses. He criticized the idea of giving AI tools unrestricted access to sensitive systems like password managers, arguing that allowing an agent to handle finance, accounts, or private vaults is “totally insane.” Stamos’ comments underline a broader industry tension over control, trust, and what users actually consent to when they enable AI assistants.
Meta maintains Muse runs in an “isolated virtual machine” to limit cross-agent or external access. The company also says many of the issues reported by staff are now being addressed, especially bugs involving confusing prompts and unintended permission use. But questions remain about how these bugs got through in a product rolled out so broadly.
Why this matters:Muse represents how mainstream AI agents are becoming, embedding in daily workflows, devices, and sensitive domains like communication and finance. If agents can’t reliably isolate private data or predictably honor permissions, they risk undermining user trust — not just at Meta, but across all AI agents. Going forward, we’ll need clearer frameworks for permission management, faster bugs-to-fix cycles, and more transparency in how and why AI assistants do what they do — especially when private data is involved.