3D Prompt Injection Defense Field

Real LLM applications pull content from many external sources at once — retrieved documents, web pages, tool responses — and any one of them can be seeded with an instruction designed to hijack the model's behavior. This 3D field simulation renders that as six orbiting sources continuously firing content chunks toward a central model core, each one first crossing a rotating classifier ring that can block it by risk, then falling toward a system-prompt priority shield that only an unusually high-weight injected instruction can breach. Unlike the 2D lane version of this simulator, here the defense is spatial: chunks travel real 3D trajectories, get destroyed at the classifier ring or physically deflect off the shield's surface when contained, and only a genuine breach travels all the way to the core and triggers a hijack flash. Tune injection rate, filter sensitivity, shield priority and sanitization to see how the defended volume holds up.