Meta’s Muse AI agent presents significant privacy risks that currently outweigh its usefulness, according to a hands-on test by a security-focused writer. The tool, positioned as an autonomous executive assistant, demands extensive access to connected accounts, including email, while retaining interaction data even after deletion requests. Meta insists its training data is anonymized, yet real-world vulnerabilities expose serious flaws in that claim.
The AI’s core function involves managing complex tasks such as travel arrangements or purchases, which requires broad permissions by design. By default, Meta uses all user interactions with Muse to refine its models, and this data retention persists even after users delete specific conversations. The company asserts that personally identifiable details are removed from training datasets, but the method is not foolproof. Users can review what Muse stores about them, though the system may still retain related information without full transparency.
Security weaknesses have already emerged. A zero-day vulnerability surfaced weeks after Muse’s launch, allowing macOS applications and terminal commands to interact with the AI and its linked services. In another case, misconfigured macOS settings granted Muse access to private messages in other applications.
Read Also: AI adoption outpaces data infrastructure development
Meta’s proposed solution, Muse Confidential VM, aims to isolate each AI agent in a virtual machine environment. While encryption would theoretically prevent even Meta from accessing user data, virtual machine containment is not inherently secure. Past breaches in similar systems from competitors like Anthropic and OpenAI show that isolation measures can fail under pressure. The writer, who specializes in cybersecurity, advises avoiding Muse on macOS entirely and recommends steering clear of the service until stronger protections are implemented.
For most users, the potential drawbacks of Muse currently surpass its advantages. Though still in beta, the service is marketed as ready for daily use, yet few individuals will examine its privacy policies or track cybersecurity developments. Many could remain unaware of vulnerabilities until an exploit occurs. Holding off on adoption until Muse Confidential VM is fully developed, or choosing alternatives with stricter data controls, may be the safer path forward.
