Bordair vs Rebuff: Prompt Injection Detection Compared

Rebuff is an open-source prompt injection detection framework that uses a multi-layered approach including heuristics, an LLM-based classifier, and a vector similarity store. It is an interesting research project, but there are important differences if you are looking for a production detection tool.

How Rebuff works

Rebuff combines four detection methods: heuristic checks, an LLM-based classifier (calling GPT or similar), a vector database of known attacks, and a canary token system. The idea is that layering multiple approaches improves catch rates.

The LLM-in-the-loop problem

Rebuff's most powerful detector is its LLM classifier, but this creates a circular dependency: you are using an LLM to detect attacks on an LLM. This adds significant latency (often 1-3 seconds per check), costs money per API call, and introduces its own prompt injection attack surface.

Bordair uses a purpose-built DeBERTa classifier, not a general-purpose LLM. Our detection runs in under 50ms with no recursive vulnerability.

Multimodal support

Rebuff is text-only. Bordair scans text, images, documents, and audio natively, catching cross-modal attacks that text-only tools miss entirely.

Maintenance and updates

Rebuff's GitHub repository has seen limited recent activity. The attack landscape evolves quickly, and a detection tool needs continuous updates to remain effective. Bordair is actively maintained with regular model updates and new attack pattern additions.

When to choose Rebuff

• You want to experiment with different detection approaches
• You are comfortable with the latency and cost of LLM-based detection
• You want a self-hosted, open-source solution

When to choose Bordair

• You need production-grade latency (under 50ms)
• You want multimodal detection
• You want a managed service with no infrastructure overhead
• You want detection that does not rely on calling another LLM

Get started with Bordair and add prompt injection protection in minutes.