Rebuff AI
Definition
Open-source prompt injection detection using multiple methods: heuristic analysis, LLM-based classification, and vector similarity. Multi-method detection for catching adversarial inputs. Stops injected prompts but does not stop harmful actions from non-injected prompts.
Why It Matters
Detection is important but assumes harmful actions only come from adversarial inputs. Models hallucinate, invent parameters, and propose mutations without any injection. Exogram's threat model includes the model's own errors — not just external attacks.
How Exogram Addresses This
Exogram matches all tool calls against threat patterns regardless of input quality. The model can generate harmful actions without being injected. Exogram catches what injection defense misses: the model's own errors, hallucinations, and schema inventions.
Related Terms
Key Takeaways
- → This concept is part of the broader AI governance landscape
- → Production AI requires multiple layers of protection
- → Deterministic enforcement provides zero-error-rate guarantees