Ground Truth.
AI, checked against the source.

News · 2026-11-17

Windows telemetry report finds coding-agent names in diagnostic classification

A reverse engineer reports that a Windows 11 diagnostic component identifies names associated with coding agents and records category bits for optional telemetry. The author observed the event in local storage and an upload queue, but did not establish that Microsoft received that particular event. Even a small usage signal can matter to developers and organizations evaluating privacy.

Key facts

The report traces a component called MissionControlAggregator in Windows’ Connected User Experiences and Telemetry service. For desktop applications, it consumes application-state and interactivity events; for Windows Subsystem for Linux, it observes execution events. When a name matches, the component sets a bit in an event called Segments.SubCohorts. The author describes an approximately daily collection window.

A bit is a compact yes-or-no marker. It is closer to a checklist saying that a matching category appeared than a transcript of what a developer typed. The author says the underlying source events can contain richer runtime and interaction details, but this component appears to retain the category bits. The published analysis does not show prompts, source code, or the number of times a tool was used in this particular event.

There is also a distinction between a name and an authenticated product. The author copied cmd.exe to a file named codex.exe; running it set the corresponding bit. That experiment supports executable-name matching, not proof that Windows checked whether the real agent ran or observed an AI interaction. A mislabeled executable can therefore create a misleading signal.

The author found the event persisted locally and placed in the diagnostic upload queue with metadata identifying it as not required. Queueing is not proof of server receipt. Microsoft’s diagnostic-data documentation separately describes optional application activity data, including launches and duration; it does not document this particular event. The report is a first-person reverse-engineering finding, and the dossier found no independent replication.

Microsoft distinguishes required and optional diagnostic data. Users of unmanaged devices can choose whether to send optional data, while organizations can apply policy controls. Microsoft’s documentation reserves a full diagnostic-data-off policy value for certain editions. Turning off optional data should therefore not be described as stopping every form of Windows telemetry.

The Hacker News discussion shows opposing interpretations: some see application-use data as ordinary optional diagnostics; others object to tracking third-party development tools or question representativeness. Both the scope and the limits matter. An organization may regard even a category bit as sensitive, while the evidence does not support a claim that Windows captured conversations or uploaded source files.

The author also distinguishes the relatively compact event from the richer source events it consumes. Windows’ underlying application and interaction events can contain details such as runtime, while the described aggregator appears to retain category bits. Microsoft independently documents that optional diagnostics can include application activity, but its general documentation is not confirmation of this specific event’s design. Conflating those two sources would imply Microsoft has documented more than it has.

For administrators, the operational question is not only whether an event is labeled optional. It is which diagnostic-data setting applies to a given edition and managed device, whether organizational policy prevents users from raising the level, and whether separately installed applications collect their own telemetry. Microsoft’s policy reference says the diagnostic-data-off policy value is limited to Enterprise, Education and Server, while Pro supports a required-data level.

There is a legitimate counterargument: diagnostic usage categories can help understand compatibility and product adoption, and users can choose to disable optional data on unmanaged devices. The privacy objection is that an indicator of a named developer tool can still disclose something about a person’s work, especially when combined with other account or device context. The report does not establish how Microsoft would use or combine this particular event, so those broader implications remain questions for the company rather than findings of the test. The observed event is a starting point for asking Microsoft for documentation, not a substitute for confirmation across builds and settings.


Primary source, verified: read the paper →

Key questions

Does the event contain coding prompts or source code?

The reported event contains category bits, not prompts, source code or usage counts.

Did the investigation prove Microsoft received the event?

No. The author observed local storage and the diagnostic upload queue, not receipt by Microsoft’s servers.
Cite this

APA

Ground Truth. (2026, November 17). Windows telemetry report finds coding-agent names in diagnostic classification. Ground Truth. https://groundtruth.day/news/windows-diagnostic-data-coding-agent-detection.html

BibTeX

@misc{groundtruth:windows-diagnostic-data-coding-agent-detection,
  title  = {Windows telemetry report finds coding-agent names in diagnostic classification},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {nov},
  url    = {https://groundtruth.day/news/windows-diagnostic-data-coding-agent-detection.html}
}

Topics: cybersecurity · privacy · telemetry · developer-tools

Comments are replies to this story on Bluesky — reply with any Bluesky account to join in.