Proton’s Lumo goes desktop, and the permission model is the real story

Giving an AI agent access to your file system and a shell is the riskiest thing you can do with a chatbot. Proton wants you to do it anyway, and it wants you to trust its version more than the ones from OpenAI, Anthropic or Meta.

The company launched Lumo Desktop for Windows, macOS and Linux on Tuesday. It lets the assistant run local tasks involving your files, connected apps and automations. That puts Lumo in the same category as the big model makers’ desktop tools, which already ask for broad access to your machine.

The timing matters. Consumer AI is moving from answering questions to acting on your behalf, and acting requires access. Apple said this week it is tightening security settings for apps that request Full Disk Access, which AI desktop apps do. When the platform owner starts adding friction, that tells you where the risk sits.

Proton’s answer is granular permissions. Each connector, whether file-system access, shell execution, browser fetching, Proton Mail, Confluence, Dropbox or Jira, can be set per action to automatic, approval-required or blocked. You can switch individual tools off mid-conversation. Scheduled tasks get their own permission sets. Connectors also support local and remote MCP servers, the standard agents use to reach outside data.

That design is sensible. Default-deny with explicit approval is how this should work. But it only protects you if people actually read the prompts, and approval fatigue is real. Set a shell connector to automatic because the clicking got annoying, and the model’s judgment is your only safeguard. Prompt injection from a fetched web page or a Jira ticket is the obvious attack path, and Proton doesn’t say how Lumo defends against it.

On data handling, the claims are stronger. Proton says chat history is encrypted so only you can read it, there are no server-side logs, and the limited data it keeps falls under Swiss privacy law. Hosting is on European infrastructure, which it says sits outside the U.S. CLOUD Act. It also says your conversations and connected-app content never feed ads or model training.

Those are promises, and some are hard to verify from outside. “Only you can read it” is difficult to square with a cloud model that has to process your prompts in plaintext at some point. Proton hasn’t explained that tradeoff here, and it’s the question I’d press first. The CLOUD Act argument is also a legal position, not a guarantee.

Proton is also building its own capacity, with a first Nvidia B300 cluster on Open Compute Project standards now going live in Switzerland. Running your own inference is the credible way to avoid shipping user data to a third-party model provider.

Lumo also adds Skills, reusable instructions, templates and reference files. Earlier Custom Lumos will migrate to that format.

The pitch is good. The proof will come from independent audits and from how the agent behaves when something hostile lands in its inputs.