Xusheng's reverse-engineering of MpClient.dll uncovered a hard-coded allowlist that singles out direct Microsoft competitors like Cursor, Claude Code, Windsurf, and Continue alongside first-party tools. He argues the specificity of the list, combined with the fact that events fire even with Defender real-time protection off and telemetry set to 'Required,' points to market surveillance rather than genuine endpoint defense.
By submitting the piece with the framing 'Windows 11 Is Collecting Data on Which Coding Agents and Developer Tools You Use,' cwaffles amplifies the competitive-intelligence read. The 384-point score suggests the HN audience largely agreed that a curated list of AI coding rivals is not what a neutral antimalware reputation feed looks like.
A portion of the 241-comment thread argues that any serious endpoint protection product profiles process launches, and that tools like Ollama which pull multi-gigabyte model weights from arbitrary URLs are exactly the kind of binaries a reputation feed should track. In this view, hard-coding names of popular new developer tools is just how Defender keeps its allowlist fresh, not a competitive signal.
Xusheng emphasizes that the telemetry fires even at the 'Required' minimum tier, and that only Enterprise and Education SKUs expose the Group Policy knob (AllowTelemetry=0) that actually stops DiagTrack. Regardless of intent, this makes the collection non-consensual for the vast majority of developers running Windows 11 Home or Pro.
On November 16, security researcher Xusheng published a reverse-engineering writeup showing that Windows 11 is quietly collecting telemetry on which AI coding agents and developer tools run on the machine. The relevant component lives inside `%ProgramFiles%\Windows Defender\MpClient.dll` and a sibling set of diagnostic DLLs that feed the Connected User Experiences and Telemetry service (`DiagTrack`). The DLL carries a hard-coded allowlist of process names — the usual Microsoft-first lineup (`code.exe`, `devenv.exe`, `powershell.exe`), but also a notably specific set of competitor tools: `cursor.exe`, `claude.exe`, `claude-code`, `windsurf.exe`, `ollama.exe`, `lm-studio.exe`, `gh-copilot`, `aider`, and `continue`.
When one of these processes launches, Windows emits a diagnostic event tagged with the binary name, a hashed machine identifier, and timing metadata — then ships it to `*.events.data.microsoft.com` over the standard telemetry pipe. The author traced the outbound traffic with Wireshark and confirmed the events fire even when Windows Defender real-time protection is disabled and even when the user has set diagnostic data to the "Required" (minimum) tier. On Home and Pro SKUs there is no UI toggle that disables the Required tier; only Enterprise and Education SKUs expose the Group Policy knob (`AllowTelemetry=0`) that actually stops the service.
Microsoft has not responded publicly as of this writing. The Hacker News thread hit 384 points within hours, with a predictable split between "this is routine antimalware behavior" and "this is competitive intelligence dressed as security telemetry."
The charitable read is that Microsoft Defender has always profiled process launches — that's how any EDR works. A hard-coded list of AI coding agents is, in that frame, just a reputation feed: Defender wants to know when `ollama.exe` starts because `ollama.exe` pulls multi-gigabyte model weights from arbitrary URLs, and that's the kind of behavior malware mimics. But the uncharitable read is harder to dismiss: the allowlist is almost a perfect enumeration of Microsoft's AI coding competitors, and the data egresses to the general telemetry endpoint, not the Defender cloud-protection endpoint. Those are different pipelines with different retention, different access controls, and — critically — different downstream consumers inside Microsoft.
This is not the first time Microsoft has been caught instrumenting developer behavior at the OS layer. In 2023, researchers found that the Windows Terminal preview was sending telemetry on command invocations. In 2024, GitHub quietly amended its Copilot privacy notice to clarify that prompt data "may be used to improve Microsoft products," which was widely read as covering Bing, Edge, and — now obviously — the Windows AI stack. The pattern is consistent: Microsoft collects developer signal at every layer it controls, then uses that signal to position its own products. Cursor's growth curve, Claude Code's adoption, Ollama's local-model footprint — these are the exact data points a product manager at the Windows AI division would pay to have.
Community reaction on HN leaned skeptical of the "it's just Defender" defense. The top-voted comment pointed out that if the goal were genuinely antimalware reputation lookup, the allowlist would be hashed binaries or Authenticode signers, not literal process names — which are trivial to spoof and therefore useless as a security signal but perfectly adequate as a market-intelligence signal. A second highly-upvoted thread observed that `aider` and `continue` are open-source Python packages with no stable exe name on Windows; their inclusion only makes sense if someone on the Windows team specifically asked for "which AI coding tools are developers running."
There is also a GDPR angle. Under Article 6 of GDPR, Microsoft needs a lawful basis to collect per-process launch data from EU developers, and "legitimate interest" is substantially harder to argue when the data egress targets a general telemetry endpoint shared with product teams rather than an isolated security pipeline. The German Federal Office for Information Security (BSI) has flagged `DiagTrack` in prior Windows 10 reviews; expect a fresh round of inquiries.
If you develop on Windows, assume Microsoft knows which AI coding agent you use, how often you launch it, and — via correlation with machine ID — can tie that back to your Microsoft account if you're signed in. There are three practical responses, in increasing order of effort.
The cheap option: run your coding agents inside WSL2. Process launches inside the Linux kernel do not surface to `DiagTrack` in the same way — the telemetry component hooks the Windows process creation path, and WSL processes appear as a single `wsl.exe` or `vmmemWSL` to that layer. This isn't airtight (WSL network traffic still traverses the host), but it strips the per-tool signal.
The real option: on Pro SKUs, apply the `AllowTelemetry` registry key via `HKLM\SOFTWARE\Policies\Microsoft\Windows\DataCollection` set to `0`, then disable the `DiagTrack` service directly (`sc stop DiagTrack && sc config DiagTrack start=disabled`). Microsoft documents this as "Security" tier and says it's Enterprise-only, but the registry key is honored on Pro in practice. On Home, you can't fully disable Required telemetry without third-party tools like O&O ShutUp10++ — which is itself a signal about how much developer autonomy Microsoft actually ships on consumer SKUs.
The nuclear option: develop on Linux or macOS. For teams that have been debating a cross-platform mandate, this is the kind of story that ends the debate. Expect renewed internal pressure at firms with any kind of competitive-intelligence concern — fintech, biotech, anyone building AI tooling that competes with GitHub Copilot.
The immediate question is whether Microsoft publishes a clarifying technical note or quietly removes the allowlist in the next cumulative update. The precedent from prior telemetry controversies says: expect a blog post framing this as "security reputation data," expect the list to be renamed but not removed, and expect no SKU-level change to who can actually opt out. The deeper question is whether OS-layer telemetry on developer tooling becomes a serious procurement gate at engineering orgs that care — and on current evidence, it should.
For anyone on the fence about switching from Windows to Linux: it takes about a week, if that, of adjusting. If most of what you do is on the Web already (and the services you use provide a Web-based client), you have nothing to lose but your chains.Take your pick - Ubuntu, Fedora, someone else ment
So does Intel, AMD, and NVIDIA drivers; they collect what executable name and paths you run; unless you opt out.Most Chrome configs end up persisting full browser history and clickstreams inside a Google account.If you use Debian and opt-in to popularity contest; it may collect and send similar info
Windows has been very slow to make a sandbox for AI agents - they launched MXC just a month ago (Oct 7) and Claude Code doesn't even support it yet. So we have people running un-sandboxed agents on Windows, or they use WSL instead. All of this to say, yes I hope they have some non-invasive tele
https://linuxmint-installation-guide.readthedocs.io/
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.
Just as AI has advanced massively in the past two years, I think a lot of gamers would be shocked at how viable Linux is as a genuine gaming platform in 2026. I already have plans to switch from Windows to Linux, I even have Linux running from an external drive to play with it.The vast vast majority