Standard AI guardrails act like a dictionary of forbidden words. They parse the semantic meaning of an incoming prompt and block it if it matches a "bad" pattern. Adversarial attackers easily bypass these systems using encoding, roleplay, or logical misdirection to obfuscate their intent from the semantic filter.
The Geometric Defense
Oblique Guard operates on entirely different principles. Instead of analyzing text, it analyzes the geometric trajectory of the prompt as it is mapped into the model's high-dimensional latent space. Malicious intents, regardless of how they are phrased or obfuscated, cluster in specific topological manifolds.
We establish "oblique hyperplanes"—strict mathematical boundaries cutting through the latent space. If a prompt's embedding trajectory intersects with or crosses these geometric boundaries, the model's forward pass is mathematically halted. By defending the geometry rather than the semantics, Oblique Guard renders complex prompt-injection attacks computationally irrelevant.