Meta is pushing hard into autonomous action with Muse, a new AI agent that lives inside WhatsApp and actually does things for you. This isn’t just a chatbot that summarizes text; it books travel, fills out web forms, and even negotiates bills or car prices on your behalf. The company claims Muse can handle complex, multi-step goals like planning a dinner party while accounting for friends’ allergies, pulling data directly from your Instagram saves to build a shopping list. It keeps running in the background, checking in only when it needs approval to send an email or complete a purchase.
The most striking move here is the integration of direct payments, a feature OpenAI recently abandoned for ChatGPT. Muse uses Stripe’s Link service to generate one-time virtual cards, keeping your actual financial details hidden from the agent itself. Meta positions this as more secure than traditional methods, offering purchase protection for damaged items or price drops. While OpenAI stepped back from handling checkout, Meta is doubling down, aiming to make the transaction seamless within the chat interface. Shop Pay and 1Password integrations are also in the pipeline, suggesting Meta wants to be the central hub for both discovery and execution.
Security is the obvious elephant in the room. Meta insists Muse runs in a “walled-off” virtual machine, monitored by a separate oversight agent called Sentinel. According to the company, Muse never sees your passwords or payment info; it only interacts with a secure store. However, history suggests these sandboxes aren’t impenetrable. Researchers have previously hijacked agentic systems using doctored calendar invites, so the claim that “nothing reaches the internet unless Sentinel clears it” will face intense scrutiny from the security community. Meta admits it hasn’t released robustness numbers, leaving some doubt about real-world resilience.
Under the hood, the intelligence is improving fast. Muse Spark scored just 31 points on the Artificial Analysis Intelligence Index in April, but version 1.3 now hits 48 on the max tier, closing in on GPT-5.6 Sol’s score of 47. Mark Zuckerberg calls this “personal superintelligence,” framing Muse as the first step toward an AI that truly knows you. Yet, while Muse data is supposedly walled off from ads, Meta still uses general Meta AI interactions for personalization. The distinction between a private assistant and a data-harvesting engine remains thin, even with promises of a future confidential VM.
