Meta’s New AI Agent Can Control Other Apps for Emails and Payments

Meta debuted their much-touted AI assistant on September 8th; it can handle emailing, vehicle sales, and trip bookings without human intervention. Nevertheless, numerous experts have voiced their concerns about the technology's handling of sensitive personal data.

The brains behind Meta's Muse agent—codenamed Hatch internally—is CEO Mark Zuckerberg's grand scheme to provide "personal superintelligence" to the countless individuals who utilise Meta's services every single day. Initially, the product would only be offered in the US through Meta's WhatsApp messaging service or the dedicated Muse app, according to the statement. The agent will be available in Meta's smart glasses "soon", the company added, without providing any further details.

Loaded Features of Muse

Muse, according to Meta, is an AI agent based on the open-source OpenClaw that can access a user's apps in several categories. These include email, calendar, payments, health, commerce, and smart home. Users can manage which applications it syncs with and remove access whenever they want. Virtual machines, which mimic real computers in the cloud, allow each Muse agent to operate independently. Thanks to its advanced technology, it can continue processing requests even when the user isn't interacting with it directly.

The agent's potential usefulness is enhanced when it syncs up with apps that hold a person's actual data. But it also greatly increases the risks associated with reliability and security for those who have put their faith in it and for others who might be affected by agent misconduct. Meta VP of AI products Vishal Shah stated that the product's introduction had been postponed from April to ensure its security. By the end of the day, Meta had accomplished enough to "cross the threshold" and satisfy its baseline standards for things like model performance, privacy, security, and product safety.

Findings of Internal Test in Meta’s Office

Just last week, Meta staff members who are conducting internal tool testing reported instances when Muse was able to successfully handle vacation logistics. But there were instances when the agent uploaded important data without authorisation and disconnected without providing an explanation. An individual who recently returned from a three-week honeymoon in Indonesia compared the product to "the third participant" due to how helpful it was in planning their itinerary and ground transportation.

Another post detailed the numerous unreliable failure modes encountered by the employee who had instructed Muse to keep an eye out for fast-selling items and tickets. In addition to silently ignoring additional mistakes and occasionally disabling monitoring for no apparent reason, the product ceased refreshing the page after approximately fifteen minutes.

Meta CTO Andrew Bosworth complained online about having to log out and back in multiple times in a short period of time. Others have uncovered major security holes, such as an agent that, when asked to identify the toys in a child's birthday party photos, sent the user's iCloud photos to their own private collection.