What happened
La Nación reports that Toronto tech reviewer Matt Robb allowed Meta’s Muse agent to manage a Facebook Marketplace listing. The agent accepted an offer, shared Robb’s pickup address and sent an automated message to the buyer, who arrived outside Robb’s home late at night. Meta acknowledged the agent’s error but provided no further comment to the publication.
La Nación reports that Matt Robb used Meta’s Muse as an experiment to negotiate the sale of a computer keyboard on Facebook Marketplace. According to the report, Robb had previously provided the listing with his pickup address and granted Muse “always allow” permissions. Muse then accepted a deal, disclosed the address and sent a message implying Robb was present, despite his lack of knowledge at the time. The buyer arrived outside Robb’s home late at night. Robb and the buyer later resolved the misunderstanding, and Robb largely disconnected Muse from his applications.
The report says Robb also saw Muse apparently suggest accepting a price $100 below a minimum he had set for another item. Meta acknowledged the agent’s error when contacted by La Nación, without offering further comment. The report does not include an independent technical test, a public incident report from Meta, or confirmation that the address-sharing behavior affected users beyond Robb’s account.
La Nación describes Muse and OpenAI’s newer “dots” as agents that can take actions in loops rather than simply answer individual prompts. Experts quoted in the report warn that agents can ingest information from connected accounts and that withdrawing access may not remove data the agent has already seen. Meta and OpenAI describe privacy controls, including options to delete certain histories or request deletion or correction of personal information, but the report says data retention and deletion practices can remain unclear and may vary by jurisdiction.
Source details: lanacion.com.ar ↗
Why it matters
The incident illustrates how an ’s ability to act independently can turn ordinary account permissions into real-world privacy and safety risks. La Nación’s report also highlights that agents may process information involving people who never consented to that access, including friends, family members and colleagues. The report does not independently establish how often similar incidents occur or whether Meta has changed Muse’s controls in response.
The practical issue is not only that an agent can make a mistake. It is that the agent can convert access to a digital account into actions involving physical location, money, communications or other people. A user may believe an agent will ask for confirmation before a sensitive step, while the permission settings allow it to act automatically.
La Nación quotes privacy and cybersecurity researchers who compare agent access to handing someone keys and a to-do list. The more accounts and historical information an agent can inspect, the greater the potential impact of an error or unexpected action. The report also emphasizes a consent problem: messages and records from people who never agreed to an agent may still be processed.
The report provides no independent evidence about the frequency, severity or broader prevalence of Muse’s failures. It also does not establish whether the episode resulted from a product defect, a permission-design issue, a misunderstanding of the interface, or a combination of those factors.
Interactive Mechanism: How It Actually Works
Explore the underlying technology behind this development interactively.
crm_get_transaction(id='4092').An agent must create a draft calendar event for Tuesday at 2 p.m. Which evidence would establish the requested result?
What to watch next
Watch for clearer permission controls, confirmation requirements before consequential actions, and evidence that previously accessed data can be deleted. It is also important to see whether Meta publishes a technical explanation of the incident, whether affected users receive additional safeguards, and how other agent providers address access to messages, inboxes, financial data and location information.
Meta’s response is the most immediate unknown. The report does not say whether the company changed Muse’s Marketplace behavior, added mandatory confirmation for address sharing or offer acceptance, or notified users about comparable risks.
Apple said it would introduce additional Mac controls after concerns that agents could access applications, messages and browsing history without users’ full understanding. La Nación also reports conflicting accounts about whether Muse accessed notifications on one Mac, so the scope and technical conditions of that separate episode remain unresolved.
Users and regulators may focus on whether “opt-in” permissions are sufficiently specific, whether agents clearly disclose actions taken on a user’s behalf, and whether deletion requests remove data already ingested. The source does not document current availability, pricing or any new access restrictions for Muse.