Guardrails & Safety
Exogram + Rebuff AI
“Prompt injection detection. Same execution gap.”
What Rebuff AI Does
- •Rebuff detects prompt injection using heuristic, LLM-based, and vector similarity methods.
- •Open-source. Multi-method detection for catching adversarial inputs.
- •Stops injected prompts. Does not stop harmful actions from non-injected prompts.
What Exogram Does
- ▸Exogram operates downstream from injection. Even clean prompts can produce dangerous tool calls.
- ▸The model can make unwarranted inferences, invent parameters, and propose mutations without any injection.
- ▸Exogram catches what injection defense misses: the model's own errors.
Key Differences
| Dimension | Rebuff AI | Exogram |
|---|---|---|
| Threat Model | Adversarial inputs | All tool calls (adversarial or not) |
| Non-Injected Threats | Not covered | Fully covered |
The Verdict
Use Rebuff for injection detection. Use Exogram because the model can generate harmful actions without being injected.
Interception Latency Benchmark
Unlike standard LLM-based guardrails that invoke external APIs on every single tool execution, Exogram compiles policies to execute in-memory inside the client runtime.
Rebuff AI (LLM / API Check)~220.00 ms
Exogram Authority Runtime (In-Memory Intercept)0.07 ms
Is Rebuff AI vulnerable to execution drift?
Run a static analysis on your LLM pipeline below.
STATIC ANALYSIS
Frequently Asked Questions
Is prompt injection the only threat to AI agents?
No. Models make unwarranted inferences schemas, invent parameters, drift constraints, and propose destructive mutations — all without any injection. Exogram covers the full threat surface.