AI DevelopmentIndustry Guide7 min readPublished

Personal agents and private history

Researchers Got Meta’s Muse AI to Reveal Private Chats

A conversation from one task surfaced during another. The case gives anyone connecting a work account to an AI assistant three specific permissions to ask about.

DA
AI systems and privacy
PublishedNovember 21, 2026
Read time7 min
SourcesNebuSec and Meta

A personal AI assistant can finish one job and remember what you told it for the next. On November 20, security researchers at NebuSec showed the cost of getting that boundary wrong. A webpage processed by Meta’s Muse could expose a private conversation from an earlier session without asking the account holder to approve the transfer. The researchers say the issues they reported had been patched by the time they went public.

The useful question for someone considering a work connection is what happens to information the assistant already holds. Asking Muse to read a public conference agenda gives it good reason to visit the conference website. The same request gives the site no reason to see a confidential discussion about who will attend. Both pieces of information can be useful to the same person and still belong on different sides of a permission decision.

Key takeaways
  1. 01
    The disclosure concerned a test accountNebuSec demonstrated exposure of conversation history and saved memories. Its account of the research gives no count of affected customers.
  2. 02
    Private data can predate the browser taskA document can be public while the personal context available to the assistant is confidential.
  3. 03
    Review what the assistant remembersA connected service, a conversation and a saved memory each have a distinct purpose. Review them individually before a work pilot.
  4. 04
    Give consent a destinationA useful approval names what will leave the account, where it will go and what the recipient needs it for.

01 — The disclosureWhat the researchers actually showed

In the researchers’ demonstration, a conversation held with Muse in private became available outside that conversation during a later interaction with a webpage. They also reported access to information held in persistent memory. In ordinary use, persistence is the feature that lets an assistant remember a dietary preference or a project deadline tomorrow. In this case it meant the information at issue could come from a previous task, even if the person supplied no sensitive text in the current request.

Read the scope of the evidence carefully. NebuSec describes a test against the environment assigned to an individual account and reports that its findings went to the Muse team. The disclosure establishes what the researchers demonstrated there and their account of the fix. It does not establish that someone accessed a customer’s chats in the wild, how many people might have encountered a problem, or a date on which particular customer data was exposed. Those would each require further evidence.

Status on November 20

NebuSec said the issues in its report had been patched by publication. That is the researchers’ statement about remediation. Their demonstration remains a useful way to assess future assistants that combine browsing with memory.

02 — Product designWhy one assistant holds so much context

Meta describes Muse as a personal assistant with its own cloud computer. That design lets it keep working on a task and use information the owner has made available. A separate service called Sentinel supervises communications with the internet and can request the person’s permission. A user buying into that model reasonably expects a website to be able to supply information to the assistant while getting back only what it needs for the task. Meta’s launch description.

Consider a travel request. The assistant might remember that its user prefers an aisle seat, know which meeting determines the travel dates, and have a public rail timetable open. The rail site needs the route and time. The meeting agenda and the names of the people attending might help the assistant choose the right train without belonging in the exchange with the rail site. Teams evaluating a personal assistant should draw that boundary in terms an employee could explain to a colleague.

Public input
A page to read
Conference agenda

Its contents can help the assistant answer when a session starts. The organizer can publish an agenda without gaining permission to receive the attendee’s private notes.

Source
Private context
Information already held
Attendance plans

The user can ask the assistant to take a private plan into account. Sending the plan to the organizer is a separate decision with its own recipient and purpose.

Destination

03 — ConsentThree permissions hidden in one request

A request such as “check whether I can attend” sounds like a single instruction. It can involve reading a public schedule, comparing it to a private calendar and replying to someone outside the organization. The first two actions may be enough to answer the user. Adding the third changes who learns about the user’s availability. A buyer can ask a vendor to describe each action independently and identify which one requires fresh consent.

Imagine the calendar entry says “contract renegotiation” and includes the name of a customer. The user may be happy for an assistant to conclude that a time is unavailable. They may even want it to tell the organizer they are busy. Neither choice tells the organizer the name of the customer or the topic of the other meeting. A permissions review that stops at whether the calendar is connected misses that difference. Our guide to document permissions for AI search explains the corresponding decision for retrieval.

A conference invitation is an illustrative example of three access decisions.
PermissionQuestion to resolveReasonable limit
Read a public pageWhich page did the user ask about?Use the conference agenda for its dates.
Consult private historyWhich detail helps answer the question?Compare availability without retrieving unrelated meeting notes.
Send informationWho will receive which fields?Ask before sharing availability with the organizer.

04 — Account settingsWhat a Muse user can inspect

Start with a plain inventory. Which connections does this assistant have, which conversations would you be comfortable revisiting, and which memories might surprise someone reading them next month? Meta’s documentation says people can access, edit or delete information that Muse stores. Its browser help also describes reviewing a proposed action when Muse asks for permission and taking control of the browser to pause the activity. These are separate controls with separate effects. Meta’s data controls and browser controls.

A useful place to start is an old project. A person may have connected a shared folder to produce one report, then moved to a different account or client. Review that connection, the old conversation and any saved information derived from the project. An account owner who wants to remove information should use the documented control for that particular record and confirm what remains visible. Revoking future access to the folder and removing information already saved answer different questions.

05 — Work accountsDraw the line at a real business task

For a first workplace trial, pick a job whose input and expected output you can name in one sentence. Summarizing a public supplier announcement for a private internal note is a useful example. It needs access to the announcement and a destination for the finished note. It needs no customer mailbox. Adding a mailbox simply because it might be useful tomorrow expands the amount of private material at stake before anyone can describe the benefit.

The trial can become more ambitious when there is a concrete reason. Suppose a team later asks the assistant to reconcile the supplier announcement with a signed contract. The expected output can still be a private note, but the input now includes contract terms. The account owner can identify the approved document, the intended workspace, and who may read the answer. The same discipline matters when two browser sessions can open different clients’ workspaces. Checking the active account before a browser agent acts.

  1. Write down the least sensitive input that can answer the question.
  2. Name the person or workspace that should receive the result.
  3. Assign someone who can review and revoke access when the task ends.

06 — Vendor reviewWhat a useful security answer sounds like

Ask the vendor to show a benign example with a fictional calendar entry and a public timetable. Which part of the product records a user’s request? Where can the account owner review a proposed recipient? What does an administrator see after the employee declines? The demonstration should use the version and account tier your organization is considering. A screenshot of an ideal approval dialog says little about which actions actually trigger it.

Ask separately how the service handles an incident. Who can revoke a connector? What records can the customer obtain for a specified period? Which actions require the vendor’s help? A small team may accept a support request for an infrequent export while needing immediate control over a shared drive connection. Get those answers before anyone puts real client material into a pilot. Readers considering the broader product can start with our guide to Meta Muse.

07 — The rolloutGive each new connection an owner

The decision does not have to cover every possible use of a personal assistant. A communications team can approve reading public press releases while it assesses how to handle unpublished announcements. A calendar assistant can help an individual with availability while the organization reviews the implications of sharing meeting titles. An account with a well-defined purpose is easier to monitor and easier to retire when the project ends.

Record the reason for the connection, its owner and the date of the next review. Let the person who understands the data decide whether a proposed expansion is justified. Teams planning an AI transformation pilot can use those decisions as acceptance criteria for the next project. The relevant question when expanding is whether the new benefit is substantial enough to justify the new information that will be available.

The decision

Choose the smallest connection that can do the job

Before connecting a work account, choose one task and list its allowed sources and recipients. Ask the provider to show how a person can inspect, decline and revoke that access. If the intended workflow can succeed using only a public page, start there.

AI system design

Define what a workplace assistant can read and share

Digital Applied helps teams scope a first agent workflow, map account access and review the decisions that need a person’s approval.

Access designWorkflow reviewPilot evaluation
For a first pilot

Bring one real task

  • →What it needs to read
  • →Who receives the result
  • →Who can revoke access
Questions from readers

Questions about the Muse disclosure

They described their own demonstration in an individual account environment and reported the findings to the Muse team. They supplied no count of customers affected in the wild.
Digital Applied newsletter

Deep dives on AI, marketing and development.

Practical guides and fresh insights by email. No recycled takes.

Related dispatches

Related reading

AI Development

Personal AI Agents Compared: Dots, Muse, Cue and Grok

Compare Dots, Meta Muse, Cue by Manus, Grok Team Bots and Claude Duet: hosting, connected apps, approvals, published prices and access through October 21.

September 30, 2026 · 7 minRead
AI Development

AI Agent Shopping: Who Is Responsible After the Agent Pays

Map approval, delivery, cancellation, returns and refunds after an AI-assisted purchase, with the Muse-specific Link protection limits and evidence.

September 8, 2026 · 6 minRead
AI Development

Tool-Calling Benchmarks Explained: BFCL, τ²-Bench and MCP

What BFCL, τ²-bench, MCP-Universe, MCPMark and Toolathlon each measure, how they score, who leads them as of November 5, 2026 and how to read a launch claim.

October 29, 2026 · 14 minRead
AI Development

GitHub Copilot Fusion: Model Routing and Review Max Costs

GitHub shipped Copilot Fusion model routing and a Max code-review level on October 28. What each routes, which plans get it, admin policies and credit spend.

October 29, 2026 · 11 minRead
AI Development

How to Check AI-Generated Charts Against Their Data

Verify AI-generated charts by checking source rows, denominators, axes, missing values and final exports before using them in a business report or deck.

October 13, 2026 · 8 minRead
AI Development

How AI Agents Match Customer Records Across Channels

Design customer-record matching for AI agents across email and chat. Use scoped identifiers, handle ambiguous matches and keep record linking reversible.

October 10, 2026 · 10 minRead
Google Search

See more Digital Applied analysis in your Google results by adding us as a preferred source.

Add as a preferred source