And there are solutions around it from other providers and model routers. OpenRouter lets you explicitly say on a per-request basis you don’t want to be routed to a provider that trains on your input, for example.

If a company really believed there’s valuable IP in code, there would at a minimum, ban laptops or ban code on laptops[0], provide and maintain a centrally managed LLM gateway[1] that handles authentication, billing and model choice, have MDM to prevent you from using your own Claude login, etc.

[0]: For example when I worked at Google, the proprietary google3 codebase cannot exist on laptops because there are no tools to download it to your laptop.

[1]: For example Claude code supports having a gateway: https://code.claude.com/docs/en/llm-gateway-connect

Heck, you don't even need to mix up a personal and business AI subscription. For the vast majority of folks one day their IDE "updated" and a bunch of AI features suddenly became available. Free, no sub needed. A dev starts using them (because, why not?) and lo and behold the company's code got shipped out in an inference call and is now scheduled to be in the next training run. Oops.

Actions speak louder than words. If the company doesn't even supply and require work devices, they don't effectively consider their code to be valuable IP.

Of course the company can still consider their code to be valuable IP.

IDE updated? It’s the company IT’s job to perform testing before distributing those updates. And also their job to use whatever managed settings to disable those unmanaged AI features.

A single digit headcount startup with a "company IT" team testing program updates before distributing them??? Um, no. You just download VS Code (or whatever IDE you want) on the Macbook they give you and it updates when Microsoft pushes a public update.