Meta Launches Muse AI Agent With Privacy-First Architecture

Meta has released Muse, a personal AI agent designed to automate digital tasks while addressing the trust deficit that has long plagued the company's relationship with users.

The agent launches today on iOS, Android, and at Muse.ai, with WhatsApp integration available now and AI glasses support coming soon. Users can try Muse for free, though heavy automation requires a paid subscription plan.

Competing in the AI Agent Market

Muse represents Meta's entry into a competitive field dominated by viral AI agents like OpenClaw and Instinct. The product comes from Meta Superintelligence Labs, the unit CEO Mark Zuckerberg established approximately a year ago to close the gap with OpenAI and Anthropic. According to WIRED, which first reported these details, Meta tested the agent internally under the codename "Hatch," with employees using it to operate third-party applications and browse the web autonomously.

The agent handles tasks through natural language prompts, from sending emails to booking travel to selling cars on a user's behalf. For purchases, Muse integrates with Stripe's Link payment tool, which generates single-use card numbers to avoid exposing real financial information across the internet. Meta says Muse is the first AI agent covered by Link's purchase protection guarantees for agents.

Security Architecture Sets New Standard

Meta's defining bet with Muse centers on privacy infrastructure. Every user operates within what the company calls Secure VM—a virtual machine that isolates their activity and keeps untrusted web data separate from the agent's action-taking capabilities.

A system called Sentinel monitors all data leaving the virtual machine, either matching it against existing permissions or prompting users directly for approval. David Singleton, vice president of engineering for consumer products at Meta Superintelligence Labs, emphasized that these human-in-the-loop prompts bypass the model entirely to guard against prompt injection attacks.

While Secure VM maintains strong privacy boundaries, Meta acknowledges the architecture isn't completely locked. Company policy prohibits accessing user Muse data, but technical access remains possible. Users can opt out of having their data used for training.

Confidential VM Raises the Bar

Meta plans to introduce Confidential VM, a more advanced architecture where each virtual machine runs in a trusted execution environment with users managing their own access keys locally. Under this model, not even Meta can access a user's agent VM.

The Confidential VM development involves Moxie Marlinspike, creator of Signal and the privacy-focused AI platform Confer. WIRED reviewed an advance draft of a technical white paper describing the system. Meta will grant select security firms access to Confidential VM source code for regular audits and will publish binaries and transparency logs so users can verify their connections.

Why it matters

Personal AI agents require unprecedented access to sensitive data—emails, financial information, third-party accounts—making privacy architecture a competitive differentiator. Meta's willingness to build systems where it cannot access user data, combined with public auditing and bug bounties up to $300,000, signals that privacy guarantees may become table stakes for AI agents to achieve mainstream adoption. For a company with Meta's trust challenges, Muse represents both a technical bet and a reputational test.

Meta has added Muse to its public bug bounty program with payouts reaching $300,000 for valid vulnerabilities, including up to $130,000 for successful prompt injection attacks affecting a single user.

These details were first reported by WIRED.