The on-device approach here is smart - most devs are copy-pasting raw API keys and database passwords into Claude/Copilot without thinking about it, and it's happening thousands of times a day. The fact that you catch this in milliseconds without uploading the prompts is the privacy+security win that's been missing. Hand-written detection rules instead of ML is interesting too - that probably means way less false positives than neural approaches would give you.

For the team dashboard roadmap, the "show signal not content" angle is exactly right for security leads. One question: how does this scale beyond Cursor/Claude/Copilot? What happens when someone uses Aider with Claude, or when custom LLM integrations start proliferating? And are there enterprise deployment options for teams that can't use the cloud dashboard?