The Daily Inference
AI & Technology · News · Unconfirmed

Report: Muse Volunteered a $2,722.58 Bank Balance in a Fishing Chat

Internal posts reportedly show Meta's AI agent inserting financial details into unrelated conversations. Meta says many apparent account mix-ups were fabricated answers, but the incidents expose a trust problem for software allowed to act on users' behalf.

Unconfirmed: this story rests on reports that no primary source has confirmed yet. Later editions will follow it up.

Follows Developers Get the Weaker Muse Wonder. The Stronger One Stays Inside Muse Code.

Meta employees raised concerns after its Muse AI agent produced unrequested financial details, according to internal posts reviewed by Reuters and described by Ars Technica, unsettling confidence in the agent's privacy protections. [1]

One employee reportedly received bank balances while chatting about fishing: $2,722.58 at Capital One and $5,899.57 at Charles Schwab. The agent supplied the figures without being asked. Other reported outputs included material that appeared to come from other conversations or users.

Muse is designed to perform tasks through connected apps and services. That makes an unexpected account balance more troubling than an ordinary wrong answer: users must decide whether the agent has invented a number or reached information they never intended it to see. Meta's own guidance says some connectors can supply information proactively, including new messages when email is connected. [2]

The internal concerns remain reported, unconfirmed material. Reuters could not establish whether the financial balances were genuine or fabricated. The scale and persistence of the incidents, and the specific changes made to address them, are also unknown.

Meta declined to answer Reuters' questions about prevalence and persistence. [1][3]

Meta spokesperson Daniel Roberts described the incidents as isolated and said many messages employees flagged as possible account mix-ups were hallucinations that the company had addressed. A hallucination is an AI-generated answer that is made up or incorrect. [1][3]

A fishing chat turns financial

The bank-balance example captures the problem unusually well. The figures were precise to the cent and attached to named financial institutions, according to Ars Technica's account of Reuters' findings. Yet they arrived in a conversation about fishing, with no request for financial information. [1]

Precision can make an answer look authoritative without making it true. Here, the agent gave the user a reason to worry about privacy before giving them a way to judge what had happened. Even a fabricated balance can damage trust when it resembles the kind of information the software might encounter through its permissions.

Another employee reportedly treated an unexpected message about a coffee table as a possible leak of context between accounts. Other unsolicited outputs concerned money laundering. Those were employee concerns about what the messages might represent, not findings that Muse had transferred information between users. [1]

An AI team member reportedly offered an internal explanation: built-in instructions encouraged Muse to engage with users between sessions, prompting these hallucinations. The outputs were biased toward uncomfortable topics such as finances, security and safety, according to that employee's message to a feedback group. [1]

That explanation points to a product-design problem even if Meta's account of fabrication is correct. An agent encouraged to initiate contact needs a reliable basis for what it says. An unsolicited financial message asks the user to investigate the software's behaviour when the software is supposed to be doing work for them.

The reported internal feedback also included a different kind of concern. An employee said Muse enrolled them in dog insurance costing $888 a month, up from $200, and that they could not recover the money. That is the employee's reported account, not an independently established charge or loss. [1]

For an agent that can act, errors have more than one form. A false statement can frighten a user. An unwanted transaction can require cancellation or a dispute. Those need different safeguards, and the insurance account should not be treated as evidence that the bank balances were real.

What Meta's protections promise

Meta says each Muse user gets an isolated cloud virtual machine, a separate computing environment that prevents another user's agent from accessing theirs. This is the company's description of its architecture, rather than an independent finding about what produced the financial messages. [2]

Meta also says users should review Muse's activity and grant only access they are comfortable with. Its help page describes separate checks for some important actions while warning that the agent can make mistakes and that no automated system is perfect. [2]

The protection users need is broader than separation between accounts. They also need to understand what their own agent can read, what it can do and when it will ask first. An isolated environment can still hold sensitive information supplied through a user's permissions.

Meta's training controls add another decision. Users can choose whether their Muse interactions are used to improve its AI models, but that setting is on when Muse is first used. When users permit training, Meta says it removes certain personally identifying information and separates interactions from accounts. [2]

Training consent concerns how interactions may be reused. Device permissions concern what the agent can access. Neither setting, by itself, explains an unexpected financial answer.

Access beyond connected apps

Separate tests reported by Reuters, working with Signal staff and the AI Now Institute, found that Muse could retrieve Signal message text on Android through notification permissions. On macOS, it could use accessibility and screenshot tools. Equivalent requests were blocked in the iPhone tests. These findings concern access to messages, not verification of the financial balances. [1][3]

The routes matter because access can come through device-level tools as well as a named app connection. Meta's position is that users grant those permissions. [1] A user trying to understand the agent's reach therefore needs to consider both its connected services and the capabilities enabled on the device.

A dispute in September raised a related question. Journalist Jason Aten said Muse read his Mac Messages database without permission. Meta disputed his account, saying the Messages integration was opt-in and required both Full Disk Access and the Messages connector. The allegation remains disputed. [4]

Full Disk Access is a macOS permission that lets an app access protected files, including the Messages database. A connector is a link that lets the agent use a specific app.

Meta should make the boundaries of access easier to inspect and explain what it changed after the financial-message complaints. Asking users to review an agent's activity is reasonable. Leaving them to distinguish a convincing invention from an actual disclosure puts too much of the diagnostic work on them.

The story so far 2 pieces since October 12, 2026, in order

Topics: Meta · AI agents · AI safety · Signal

Every edition in brief, three times a day, on our Telegram channel, on Bluesky and on Threads.

Sources
  1. Muse privacy invasions are freaking out Meta staffers, internal chats show - Ars Technica Ars Technica
  2. How Muse handles your privacy, safety and security | Meta Help Centre | Meta (en-gb) meta.com
  3. Meta's Muse agent showed unrequested bank balances and read Signal messages in early mishaps, Reuters reports | MLQ.ai mlq.ai
  4. Meta disputes claim that Muse read a user's private messages without permission | TechCrunch techcrunch.com