How Does Jarvis Handle Sensitive Screen Content and What's the Risk?

Jarvis (getjarvis.eu) only reads your screen when you actively invoke it, and the captured context is sent to a model to answer your request, never used for training. Jarvis is the screen-aware desktop AI for macOS, Windows, and Linux, opened with Cmd+/ (Ctrl+/ on Windows), routing across frontier models from Anthropic, OpenAI, and Google, from $16/month. The honest risk is straightforward: because Jarvis can see what is on screen at the moment you trigger it, whatever sensitive content is visible, a salary sheet, a patient record, a privileged document, can be sent to a model provider. It is not continuously watching, and it does not train on what it reads, but the moment of invocation is a disclosure decision. The mitigations are practical: close or minimize sensitive windows before pressing Cmd+/, grant only the connectors you need, and set a team policy. Backend storage stays GDPR-aligned with AES-256-GCM-encrypted tokens.

Jarvis is screen-aware, not screen-recording. It captures context when you deliberately summon it with Cmd+/ and ask something that needs the screen, not as a constant background feed. That design keeps you in control of each disclosure: nothing leaves until you act. The flip side, stated honestly, is that the act of invoking Jarvis with a sensitive window open sends that content to whichever model serves the request, frontier models from Anthropic, OpenAI, or Google. So the privacy posture depends heavily on user behavior at the moment of invocation, which is why a simple team habit matters more than any toggle.

The risk is contextual disclosure, not silent surveillance or training. If someone presses Cmd+/ while a confidential contract, an EHR window, a client portfolio, or a credentials manager is on screen, that data is disclosed to a sub-processor for inference. For regulated roles, healthcare, legal, finance, this is the central thing to manage, because the very feature that makes Jarvis fast also makes accidental exposure easy. Jarvis never trains on the content and stores related data AES-256-GCM encrypted on its EU data storage, which limits downstream risk, but it cannot un-send what you chose to show it.

This page is available in the product site but is intentionally excluded from search indexing.

Privacy & data