AI Guardrails

Anything you feed a language model is an attack surface: a pasted email or an uploaded document can carry instructions aimed at the model rather than the reader. Two Edge Attention screens check text on your device before it would reach a model. EdgeGuard-1 looks for prompt-injection attacks, EdgeSafe-1 for hate, abuse, and profanity.

EdgeGuard-1 and EdgeSafe-1 are built on Apache-2.0 licensed components and prepared for the browser by Edge Attention; notices ship with our repository. Text is processed on your device and never uploaded.