Skip to content

Recall in an Intrusion: a Target and a Source

What attackers can take from a Recall store (infostealers, TotalRecall, TotalRecall Reloaded), what it gives responders, and the limits of a snapshot.

Published on 10 min read

TL;DR. A Recall store is a target: it holds the passwords, messages and documents that were on the user's screen, minus what its filters caught. In the 2024 pre-release builds it was a plain SQLite database in the user's profile that infostealers could copy in seconds. In released builds the files are encrypted, but Alexander Hagenah's 2026 TotalRecall Reloaded showed that code running as the user can collect decrypted content once the user unlocks Recall, and that some data can be read or deleted without Windows Hello; Microsoft says this is within its design. For responders, a readable store, or the user's own unlocked session with proper authority, shows what was on screen: the phishing email, the file name as the user saw it, commands and their output in a console, web pages, chats. It shows nothing that never reached the screen, and a snapshot proves display, not intent. The tool on this site puts readable snapshots on a timeline with full-text search of the OCR text and flags a few things worth a look.

Recall as a target

2024 pre-release builds

Kevin Beaumont and Alexander Hagenah's original TotalRecall made the point in May and June 2024:

  • The database and images were plain files under %LOCALAPPDATA%\CoreAIPlatform.00\UKP\. TotalRecall copied ukg.db and ImageStore\, then searched the OCR text for words such as "password".
  • The data was small: Beaumont measured about 90 KB for several days of work, compressed. Exfiltration takes seconds.
  • Beaumont tested an off-the-shelf infostealer. Microsoft Defender for Endpoint detected it, but automated remediation came more than ten minutes later, after the Recall data had left.
  • Deleted and disappearing messages remained in the database, and the history included every website visited.

So, on a machine with such a store, if an intruder ran code as the user, assume the store was copied.

Released builds

The reworked feature encrypts the store and gates viewing with Windows Hello (Microsoft). Microsoft's design also includes rate limiting and protections against malware that tries to "ride along" with the user's authentication. Research since then has found weaker points around the vault rather than in it:

  • A human with the PIN. Beaumont found in April 2025 that after setup, the Windows Hello PIN alone opens Recall. In his test, a non-technical person who had seen his PIN found his Signal conversations from the day before in five minutes. Shoulder surfing, an abusive partner or a colleague at an unlocked desk is a threat model to keep in mind.
  • The sensitive information filter misses things. Beaumont found a card number and its CVV saved and indexed.
  • Code running as the user. Hagenah's TotalRecall Reloaded (April 2026) does not attack the encryption, which he calls sound. It targets the process that draws the Recall timeline, AIXHost.exe, which receives the decrypted snapshots, OCR text and metadata after the user authenticates. Code running with the user's standard rights can collect what passes through it, either by prompting the user to open Recall or by waiting until they do. He also reported that the most recent snapshot, some storage details and the delete function are reachable without Windows Hello. He tested on an ARM64 build of Windows 11 (26300.8155).
  • Microsoft's answer. It closed his report on 3 April 2026, saying the behaviour "operates within the current, documented security design of Recall" (iTnews). Whether later builds changed anything is not public at the time of writing.

This site does not describe how to reproduce these methods. For a responder, the practical reading is: on a released build, if attacker code ran in the user's session and the user opened Recall afterwards, the content may have been exposed.

Policy changes

An intruder with administrator rights might re-allow Recall. SigmaHQ's rules look for reg.exe deleting DisableAIDataAnalysis or setting it to 0, and for the same changes in registry events. Saving snapshots still needs the user's opt-in, with Windows Hello. A policy change followed by a store that starts writing in the intrusion window deserves the question of who opted in, and when. See Recall policies and settings as evidence.

What to check on the target side

QuestionWhere to look
Did the compromised user have a store?CoreAIPlatform.00 in their profile, its size and times (disk image)
Could the store be read as files?The build and the file itself: a 2024 store opens as SQLite
Did something other than Recall read the store?EDR file events on CoreAIPlatform.00; archives containing ukg.db; icacls run against the UKP folder, which the 2024 TotalRecall script does
Did code reach the timeline process?EDR telemetry for remote threads or unexpected modules in AIXHost.exe; Recall opening when the user did not open it
Was the history deleted?A drop in the store's size, deletions in the USN journal, at a time that matches the intruder's activity
Which secrets were on screen?The snapshots themselves, if readable: search the OCR text for passwords, tokens, recovery codes

Rotate any credential that appears in a snapshot from the exposure period, whether or not you can prove it was taken.

Recall as a source

When a store is readable, or when the user can show their own session under proper authority, Recall answers questions that other artifacts leave open:

  • What the user saw: the sender, subject and text of a phishing email; the file name as it appeared in File Explorer, where Windows hides known extensions by default; a fake error message shown by a malicious file.
  • What was typed in a console, and the output. CyberCX found commands and their results in the OCR text of Windows Terminal snapshots, without any logging configured.
  • Web pages and URLs, including pages that left no browser history, and the window title changes of private windows, which Recall recorded in 2024 while cutting the windows out of the images.
  • Chats and calls, including deleted and disappearing messages, and live captions of calls (Beaumont).
  • Window events: which windows opened and closed, and when focus moved, between snapshots.

Remote sessions

Be precise about what Recall can see:

  • Outgoing connections from the Recall machine. Microsoft filters Remote Desktop Connection (mstsc.exe), VMConnect, the Azure Virtual Desktop MSI client and RemoteApp windows by default; other clients are saved unless they protect their windows from capture. Beaumont found Azure Virtual Desktop apps, Windows App and AnyDesk sessions in his snapshots.
  • An intruder driving the user's own desktop through a remote-control tool: Recall runs in that session and captures its screen, whoever moves the mouse. This follows from how Recall works; no published test covers an intrusion of this kind.
  • A separate logon to the machine over RDP by another account: no public source documents Recall saving snapshots of such a session. Stores are per user, and each user must opt in with Windows Hello. An account created by an intruder is unlikely to have a store; if it has one, find out how it was enrolled.

Reading snapshots carefully

LimitConsequence
A snapshot shows the screen, not an action"The email was displayed at 09:05" is supported; "she read it" is not
Snapshots are taken every few seconds, when the content changesAn event between two snapshots may not appear; the snapshot time is not the event time
Hidden windows never appearA PowerShell started without a window by a malicious file leaves no snapshot
Filters remove contentPrivate windows, DRM, filtered apps and sites, sensitive information, DLP: a gap is expected, not suspicious in itself
OCR makes mistakesSearch variants, and read the image before quoting a string
A snapshot holds the whole screenText in a background window is not necessarily what the row's title describes; check IsForeground
Stores are per userOne user's store says nothing about another user's session

Corroborate every finding with an independent source: the mail store for the email, the file system and Amcache for the file, the PCA launch log or Prefetch for the launch, the browser history for the page.

What to look for

PatternWhyBenign look-alikes
A credential-looking string in OCR textA secret was on screen and is now in the storePassword reset pages, documentation, test data
An executable name with a document-like double extension in File Explorer, a mail client or a download barThe moment a lure reached the userRare
A console window with commandsHands-on activity, with its outputAdministrators, developers
Remote access or file transfer tools on screenControl or exfiltrationSupport sessions
A download page just before an executable appears in DownloadsDelivery pathOrdinary software installs
A history that starts afresh, or a store that shrank, in the intrusion windowReset Recall, a policy deletion, or a delete call by code running as the userThe user managing their privacy; retention

In reports, mask credentials and third-party personal data, and quote only what the case needs. A walk-through of these steps on a synthetic store is in Investigating with Recall snapshots: FIN-WKS-07.

FAQ

Can malware steal Recall data on released builds?

Not by copying the files, which are encrypted. But in 2026 Alexander Hagenah showed that code running as the user can collect decrypted snapshots, OCR text and metadata from the process that displays the Recall timeline, once the user has unlocked Recall with Windows Hello. Microsoft closed his report as not a vulnerability, saying the behaviour is within Recall's documented design.

Should I treat a Recall store as exposed after an intrusion?

On a 2024 pre-release store, yes, if the intruder could run code as the user: the files were plain and small. On a released build, consider it possibly exposed if attacker code ran in the user's session while the user opened Recall. In both cases, rotate the credentials that appear in the snapshots.

Does Recall record what an intruder does over RDP?

No public source documents Recall saving snapshots of a session that someone opened on the machine over RDP. Stores are per user, and each user must opt in with Windows Hello. Microsoft filters some remote desktop clients on the Recall machine itself. An intruder who drives the user's own desktop through a remote-control tool is a different case: whatever appears in that session is what Recall captures.

Can an attacker erase the Recall history?

A user can delete snapshots or reset Recall at any time, with Windows Hello. Hagenah also reported that code running as the user can call Recall's delete function without Windows Hello. Look for a sudden drop in the store's size and for deletions in the USN journal.

Frequently asked questions

Can malware steal Recall data on released builds?

Not by copying the files, which are encrypted. But in 2026 Alexander Hagenah showed that code running as the user can collect decrypted snapshots, OCR text and metadata from the process that displays the Recall timeline, once the user has unlocked Recall with Windows Hello. Microsoft closed his report as not a vulnerability, saying the behaviour is within Recall's documented design.

Should I treat a Recall store as exposed after an intrusion?

On a 2024 pre-release store, yes, if the intruder could run code as the user: the files were plain and small. On a released build, consider it possibly exposed if attacker code ran in the user's session while the user opened Recall. In both cases, rotate the credentials that appear in the snapshots.

Does Recall record what an intruder does over RDP?

No public source documents Recall saving snapshots of a session that someone opened on the machine over RDP. Stores are per user, and each user must opt in with Windows Hello. Microsoft filters some remote desktop clients on the Recall machine itself. An intruder who drives the user's own desktop through a remote-control tool is a different case: whatever appears in that session is what Recall captures.

Can an attacker erase the Recall history?

A user can delete snapshots or reset Recall at any time, with Windows Hello. Hagenah also reported that code running as the user can call Recall's delete function without Windows Hello. Look for a sudden drop in the store's size and for deletions in the USN journal.

Related articles

A worked example on a synthetic Recall store: alice's morning, the invoice email, the .pdf.exe opened at 09:12:07, what Recall shows and what it cannot show.
Check Recall's status, copy the CoreAIPlatform.00 store with its WAL, save the policy hives, and settle authority, privilege and third-party content first.
On a disk image of a released build, Recall content is encrypted. What you can still establish: which user had Recall, when it was saving, and deletions.