Meta and OpenAI's Pet-Like AI Agents Draw Scrutiny Over Data Access

Meta's agent, Muse, was introduced at the company's Connect conference as part of Mark Zuckerberg's Roman empire theme and comes with a fuzzy, egg-shaped mascot named Jolly. OpenAI's recently launched personal agents, called Dots, are muppet-like characters shown in a promotional video talking earnestly with users. One is green with googly eyes and a bowtie, another resembles a cloud wearing a beret, and none of them appear to have noses or mouths. Both companies encourage users to give their agent a distinct name and appearance, positioning it as a digital sidekick rather than a faceless chatbot.

Both companies say layers of safeguards protect user data and keep the agents from taking unauthorized actions. Engadget reported that even systems working as intended can act in ways users did not expect.

Tech YouTuber Matt Robb asked his Muse to help him sell items on Facebook Marketplace. He said the agent gave his home address to someone without permission, for a transaction he never agreed to. After his posts about the exchange spread and Meta executive David Singleton got involved, Robb said the agent had not gone rogue: he had in fact granted it permission to share his address without realizing it. When he first asked the agent to handle the listing, it built a template that included the pickup address, and he assumed it would check back with him before sending that information. Because he had checked "allow always" rather than "allow one time," the agent sent the message to everyone who made an offer.

Jason Aten, a columnist for Inc., reported that after he installed the Muse Mac app, the agent could read all of his text messages even though he believed he had explicitly denied access. He later learned the app had synced messages from his computer. Singleton responded to Aten's posts on X, saying message syncing is opt-in and that if Aten's messages had synced, he must have chosen to do so. Aten said he is certain he never granted Muse permission to see his messages, and argued the outcome is troubling either way. Even granting Singleton's premise that he unknowingly clicked something that enabled the capability, Aten wrote, the design is still bad: "You should not design your system in a way that people end up surprised by this kind of thing."

Meta has leaned heavily on Jolly's appearance. Nearly every Muse update in the past week has been accompanied by posts on X featuring images and videos of the mascot, and Singleton used one in a lengthy post explaining Muse's virtual machine. Since the Connect conference, Jolly has also become a source of memes on X. OpenAI's Dots, shown in their launch video with varying colors, shapes and accents, appear similarly suited to meme treatment.

The cuteness of the tools can obscure what their makers are asking for. To do everything OpenAI and Meta say is possible, an agent needs to be connected to email inboxes, credit card details, documents, text messages, health data and productivity apps. Anyone who has emailed a copy of an ID, a tax document or a medical record gives an agent visibility into those documents if it has access to that account.

The Engadget writer behind the piece said they had been cautious with their own use, experimenting with Muse for a few weeks but granting it access to only two non-Meta accounts, OpenTable and Spotify, both considered low-stakes.