Latest in AI
Windows 11 Coding-Agent Telemetry: What to Check in Privacy Settings
A researcher found an optional usage report, but the evidence does not show that every Windows PC sends it.
A Windows 11 investigation points to a reason to review the operating system's diagnostic settings separately from your coding agent's privacy controls.
What the researcher actually found
In his November 16 investigation, Xusheng Li describes executable-name matching in MissionControlAggregator.dll and a Segments.SubCohorts event that records usage categories. He found the event in Windows' upload queue with optional diagnostics enabled, not merely as an unused string in a binary.
There are important limits. His ARM64 virtual machine had the coding-agent feature enabled through a service experiment, while the public-ISO baseline did not. He found no prompt text or source code in the examined report. His renamed copy of cmd.exe also triggered a Codex flag, so a flag alone does not establish the actual application used.
I have not reproduced the binary analysis or inspected a Windows machine for this article. The settings steps below come from Microsoft's documentation; the component-specific findings remain attributed to Li.
Check optional diagnostic data first
Microsoft's privacy guide gives this Windows 11 path:
- Open Settings, then Privacy & security.
- Select Diagnostics & feedback.
- Under Diagnostic data, review the optional-data choice. If you do not want optional diagnostics sent, turn that option off.
Microsoft says Windows remains secure and operates normally with only required diagnostic data. Turning off optional diagnostics does not mean all Windows diagnostic collection stops. If the control is unavailable or marked as managed by your organization, ask your administrator rather than trying to override the policy.
Record the setting you actually see and the date you checked it. An installation-default claim from another computer is less useful than the current state of your own machine, especially for a managed work device.
Inspect events without treating the viewer as an opt-out
Microsoft's Diagnostic Data Viewer guide explains how to inspect diagnostic event details. Download the viewer from the Microsoft Store link in that guide, enable data viewing under the same Diagnostics & feedback page, and open the app.
Search event details for MissionControlAggregator or Segments.SubCohorts, the identifiers in Li's report. If a matching event appears, inspect the JSON and record the relevant fields privately. An absent result is not proof that the component has never run, and Microsoft warns that a visible event may still be queued rather than already uploaded.
Data viewing can use up to 1 GB by default. Microsoft recommends switching it off when finished; that removes the viewer's Windows data history. Turning off data viewing is not the same decision as turning off optional diagnostic sending. Keep the inspection control and the privacy control separate.
Do not upload raw diagnostic exports to a public issue or chat just to ask whether a field is normal. First identify the smallest relevant excerpt and remove personal details. This is a review task, not a reason to distribute a new copy of the machine's activity records.
Deleting old data is a separate action
The Microsoft privacy guide also provides a Delete diagnostic data control. It does not stop future sending, and it does not delete diagnostic information associated with the Microsoft account. Organization-managed devices can have additional copies under IT control.
That distinction matters when someone says they have “turned telemetry off.” Ask what changed: optional sending, local viewer history, previously collected device data, or a setting inside the coding agent. Those are different actions and should not be collapsed into one checkbox in a team checklist.
Make the check repeatable
For a small engineering team, I would keep a short record of the device's diagnostic setting, whether policy manages it, and who owns the next review. If you inspect events, distinguish an observed event from a conclusion about what was transmitted. Avoid copying the researcher's service restarts or diagnostic-state modifications onto a production laptop simply to force a result.
The finding is worth taking seriously without expanding it beyond the evidence. A useful privacy review ends with a known setting and a clearly described observation, not a more alarming headline.
I'm Paras. I work at the intersection of AI, engineering and distribution.
I help developer-tool teams explain technical findings and product tradeoffs in clear, practical content.