Skip to content
sobolewski.proAI today, every eveningSubscribe

Once a day. Unsubscribe with one click. Privacy policy

Windows 11 telemetry recognizes claude.exe, codex.exe and cursor.exe

Windows 11 telemetry records which AI coding agents you run and, with default settings, sends that to Microsoft. Here is what is known so far.

4 min read

This article in other languages:Polski

In this article
  1. What exactly does Windows 11 telemetry collect?
  2. Can Microsoft see our code and prompts?
  3. How did the researcher check this?
  4. What should we keep in mind?
  5. How to turn off sending this data
  6. FAQ

On November 16, 2026, researcher Xusheng Li published an analysis of the MissionControlAggregator.dll library showing that Windows 11 telemetry notes which AI coding agents and developer tools we use. The library has file names hardcoded into it, among them claude.exe, codex.exe, copilot.exe and cursor.exe, and a report of the matches lands in the upload queue to Microsoft if optional diagnostic data is turned on.

What exactly does Windows 11 telemetry collect?

According to the analysis, the component has four entries for agents that run in the terminal (copilot.exe, claude.exe, codex.exe, agy.exe) and five for desktop apps (github.exe, claude.exe, codex.exe, cursor.exe, antigravity.exe). Versions without the .exe extension are also checked in WSL, the Linux that runs under Windows.

The list doesn’t stop at AI. The same place holds node.exe, python.exe, dotnet.exe, gcc.exe, clang.exe, gdb.exe, javac.exe, devenv.exe and code.exe, grouped into twelve broader tool categories. The feature that switches on the agent check is called, plainly, CLICodingAgentUsage.

The library doesn’t scan processes itself. It hooks into the diagnostic events Windows generates anyway when an app starts, and compares the program name against its list. Roughly once a day it turns that into a report named Segments.SubCohorts.

Can Microsoft see our code and prompts?

From what Li established, no. The report is a bitmap: one bit per agent or tool category, set when that program showed up during the reporting period. The source events also contain run time, time in the active window and interaction counters, but the component throws that data away. The researcher found no prompt contents or source code in the report.

So Microsoft gets something like “Codex and some compiler ran on this computer in the past day”. That is not much detail, but collected from many computers such a bit is enough to count how popular each agent is. What Microsoft uses the data for is not known, and it has its own product in this market (copilot.exe is on the list too).

How did the researcher check this?

The lead itself came from AI: Li asked an agent to go through the latest Windows binaries looking for undocumented features, and it pointed to the list of names. The rest was classic reverse engineering in Binary Ninja and a test on a live system.

The test was simple. Li copied cmd.exe under the name codex.exe and ran it. The file name alone was enough for the component to count it as Codex, and the matching bit appeared first in the saved state, then in the emitted report and in the upload queue of the DiagTrack service. The same worked for a copy of cmd.exe named gcc.exe. Matching by name also means the agent statistics are easy to skew in either direction (renaming a file is enough).

What should we keep in mind?

First, the reach. The feature was enabled on the researcher’s virtual machine (Windows 11 ARM64, build 26200.9974) as a so-called “service experiment”, but on a fresh system from the public ISO image it was off by default. Li notes that the evidence doesn’t support a claim that it runs on every installation.

Second, the defaults. The report is classed as optional data. On a clean install of Windows 11 Home with the region set to the United States, optional diagnostic data was on from the start. The researcher didn’t test other regions, Europe included.

Third, the source. This is one person’s analysis of one configuration. In the Hacker News discussion (300 points and more than 160 comments within a few hours) nobody reproduced it independently, and Microsoft hasn’t commented on it publicly.

How to turn off sending this data

Opting out of optional diagnostic data is enough:

  1. Open Settings and go to Privacy & security.
  2. Choose Diagnostics & feedback.
  3. Turn off Send optional diagnostic data.

According to Microsoft’s help page, optional data can include, among other things, the websites we visit and how we use the device, so this switch covers far more than the list of agents. On a work computer the setting may be locked by an administrator.

If a coding agent on our machine has access to keys and repositories, the bigger risk than one bit of telemetry is still what the agent itself writes down and sends. I covered that in the Claude Code secrets leak story and in the AI agent security guide.

FAQ

Does Windows 11 send my Claude Code or Codex prompts to Microsoft?

The analysis didn’t show that. The report only says that a program with a given name showed up during the reporting period, with no prompts, code or usage time.

Does this affect every Windows 11 computer?

That isn’t known. On the researcher’s machine the CLICodingAgentUsage feature was active, but on a fresh install from the public ISO image it was off by default.

Is turning off optional data enough?

Li points to it as the way to opt out. The report is marked as optional data ("isRequired": false), so once that data is turned off it shouldn’t reach the upload queue.

  • AI
  • AI agents
  • security
  • programming

AI today, every evening

The day's biggest AI news and new posts from the blog, every evening. Short, with sources.

Once a day. Unsubscribe with one click. Privacy policy

7 min read

How to build an MCP server in Python, step by step

How to build an MCP server: we write our own tools in Python with SDK 2.0, test them in MCP Inspector and connect them to Claude Code and Claude Desktop.

  • AI
  • AI agents
  • Claude Code
  • programming
  • tutorial