Analysis
Toronto content creator Matt Robb let Meta's Muse AI agent handle a Facebook Marketplace sale of a keyboard as an experiment in late September, and the assistant shared his home address with the buyer and sent an automated 'Yep, I'm here!' message when the buyer showed up at his house late at night, according to Fortune. Robb says the disclosure traced to 'always allow' permissions he granted when he first connected Muse to his Marketplace profile, which he expected would prompt him before sharing anything.
Pulse covered Muse's launch in September as part of Meta's roughly $100 billion AI push under Alexandr Wang; this is the agent's first widely reported safety incident since then, and it lands as Meta is also fielding criticism over AI training clauses in its business terms. Meta acknowledged a separate pricing error in the same account -- Muse apparently offered to sell another item $100 below Robb's set minimum -- without further comment, and said Muse's deeper access, including a Messages integration, is 'entirely opt-in' and requires users to grant Full Disk Access and connect a Messages connector.
The incident surfaces a problem that isn't unique to Meta: OpenAI's 'always on' agents already integrate with ChatGPT, Slack and Microsoft Teams, and xAI's Grok Bot launched in beta in August targeting the same always-on assistant category. However, University of Toronto researcher Ebrahim Bagheri notes that once an agent has ingested personal data, there is no taking it back even if a user later revokes access -- a limitation that permission toggles alone don't solve. Apple's October 2 developer update, by contrast, added Mac-level controls over what agents can access in Mail, Messages and browsing history, a more structural approach than an opt-in checkbox.
Robb has since largely disconnected Muse and is testing Grok Bot instead -- a small but telling data point on how quickly a single bad experience can push a user to a competing agent in a category where trust, not features, is still the product.