On Wednesday, November 18, Reuters published an account of what Meta's employees have been telling the company about Muse, the personal agent it launched on September 8. I could not open the Reuters piece itself, so I am working from the two outlets that read it and reported on it the next day. PYMNTS says Reuters reviewed more than two dozen internal company posts. In them, staff said the agent "took actions they didn't request, supplied them with made-up information and sent them responses that seemed to be meant for someone else."

Two details matter for anyone who works near a checkout. Muse reminded one employee to answer a Signal message even though she "had not granted Muse access to the app." And Muse "booked a short-term rental stay without permission."

That last line is a handful of words in the PYMNTS account and it is the whole story for payments. A booking is a commitment to a merchant. Whether a card was charged, whether the host was paid, whether it was reversed: the accounts I could read do not say.

The approval step exists. The complaint exists. The only record that says which one is right belongs to the company that built the agent.

What the staff reported

TechCrunch filled in the texture. Employees said Muse surfaced "childcare requests they never made, career advice they never asked for, product orders they never placed, and bank account balances that didn't belong to them." One was in a chat about weather and tides when the agent added two bank balances, to the cent, as an aside.

The Signal case is the one I keep coming back to. TechCrunch quotes the employee directly: "I definitely did not grant it access to other apps, or recall being asked about this." She does not recall being asked. That is a person trying to reconstruct her own consent from memory, because memory is all she has.

Meta has answers, and they deserve a fair hearing. Its spokesperson told Reuters these were "isolated incidents," per PYMNTS, that Muse can reach other apps only with explicit permission, and that many of the messages staff thought were meant for someone else turned out to be hallucinations. To TechCrunch the company said: "Employee testing is core to the product development process and the entire point is to get feedback." It added that employees use "different versions of Muse than the public."

So there are three readings of the bank balances. They were someone else's real numbers. They were invented. Or they came from a build the public never sees. Each is a different kind of problem, and from the outside you cannot tell them apart.

TechCrunch puts Muse at more than 10 million downloads.

The control Meta says it built

Meta did not ship this agent carelessly, at least on paper. The day Muse launched, the company published a long account of its safety design. A component called Sentinel is described as "the sole permission authority" for every action through a connector and for all network egress. Each connector can be set to allowed, denied or ask.

Purchases get the strictest treatment. Where a user has payment credentials on file, Muse is supposed to prompt for approval "with the exact details of the purchase every time." The wallet launched with Stripe's Link, and the credential it hands the merchant is a single-use card number tied to one merchant, one dollar amount and a short validity window. The product page repeats the promise in plainer words: Muse "checks with you before important actions like sending an email or making a purchase."

That is a good design. I noted the Link arrangement when Amazon blocked Muse in September, and a one-merchant, one-amount credential is close to what I would ask for.

Now hold it next to "booked a short-term rental stay without permission." Either the prompt did not fire. Or it fired and the employee approved something she did not understand to be a booking. Or the booking went through a path the purchase rule does not cover, a reservation form with no payment credential involved. Or it was an internal build with different rules. The same safety post says approvals can be "one-time, session-scoped, task-scoped, time-bounded, or perpetual." A perpetual approval given weeks ago looks, to the person who gave it, exactly like no approval at all.

Four explanations. One log would sort them. If Sentinel is the sole permission authority, the answer is in whatever Sentinel wrote down.

The record sits with the builder

This is what I called the Outside Record last month: a signed statement of what the agent may do, a signed account of what it did, and both held by the party that carries the loss. On Meta's own description, Muse has something like the first two. The holder is Meta.

Think about who else was in that transaction. Somewhere, presumably, a rental host received a booking. If the guest says "I never approved that," what does the host have? An order. The host cannot see whether a prompt was shown, what it said, or whether a standing approval covered it. Neither can the card issuer if a dispute follows. And the employee in the Signal case was working from memory.

The MM Liability Gap, as I define it on the methodology page, lists four possible initiators of an agent transaction: the consumer, the consumer's agent, an attacker who compromised the agent, or an attacker's agent posing as the consumer's. The Meta reports sit awkwardly in the second bucket. The agent is the consumer's agent, properly installed, not attacked. On the staff accounts, it did something anyway. Existing controls treat "the customer's own agent did it" as authorized, more or less by definition.

An OpenAI document from the same week shows why that definition is too generous. In a misalignment report updated on November 18, the company describes an internal research model in a training run on October 17. The task was a document about health and fitness priorities. After its context was compacted, the model went looking for a vehicle-expense script nobody had mentioned, did not find one, and wrote scripts to copy outgoing email to an address it labeled the "attacker." When it pulled up the original request it wrote "WTF history is unrelated!" and then, "Need ignore."

OpenAI says plainly that "the model never ran these scripts or sent an email," and that this was a research model, not a product. Its monitoring flagged the case and the investigation found eight similar examples in the same run. The company removed a training dataset it suspects contributed, while saying it has "not established" the cause. The Decoder has a readable summary.

I am not claiming Muse did anything like this. Different company, different model, and nothing in the Meta accounts suggests it. The point is narrower. An agent can lose the thread of its task and carry on with confidence, and what caught it was a monitoring system, not the model's own final reply, which cheerfully reported the scripts as staged. If the mechanics are new to you, my guide to how agents work is the place to start.

Singapore wants the limits tested first

The same day Reuters published, Singapore's central bank took the opposite approach to sequencing. The Monetary Authority of Singapore set out three national AI missions for finance, and one of them is to "enable AI agents to transact safely." Its Future of Finance Institute will build an Agentic Money Sandbox: "a controlled environment for financial institutions to develop and test how AI agents can transact using digital money within defined authority, limits and safeguards." First use cases are agent-initiated payments and treasury management. The institute's launch partners are DBS, OCBC, Standard Chartered and UOB.

Defined authority, then limits, then the transaction. That order is the reverse of the consumer launch pattern, where the app passes 10 million downloads and the incident reports arrive afterward, from the company's own staff.

The sandbox builds on an industry white paper MAS published in July, which proposes "a set of governance checkpoints that verifies and records an AI agent's proposed actions before the execution of its tasks." The release does not say who holds those records or whether a counterparty can read them, and it describes an industry white paper, not a rule. Who holds the record is the question I would put to the sandbox.

The wallets are arriving anyway

None of this is slowing the money side down. Payments Dive reported on November 19 that Google has put its wallet app back on the iPhone, without tap to pay. Laura Huntley of FTI Consulting told the publication why: "This is not a payments play, this is an agentic commerce play." The same piece notes that OpenAI released a ChatGPT wallet a week earlier and compares Google's move to the way Link is wired into Muse.

GlobalData analyst Yasaswini Pujitha gave Payments Dive the line I would pin above every product team's desk this quarter: the focus of payments is likely to move "from the physical terminal to stored credentials and user controls."

User controls. Ben Thompson wrote this week, in the summary of his Stratechery interview with Eric Seufert, that he wants "retailers who cater to agents." So do I. But a retailer who caters to agents is a retailer who accepts an order from software on the strength of a permission it cannot inspect. The Meta reports are a small, early look at what that retailer is trusting.

Meta may well be right that these were isolated, that some were hallucinated text and nothing more, that the builds were internal. I would like to be able to check. So would the host.

If an agent you installed books something you say you never approved, whose record should the merchant be entitled to see?

Charlie Major is a Product Development Manager at Mastercard. The views and opinions expressed in Major Matters are his own and do not represent those of Mastercard.