Bordair vs Meta PromptGuard: Which Prompt Injection Detector Should You Use?

Meta's PromptGuard is an open-source prompt injection classifier built on mDeBERTa. It is a solid research contribution, but using it in production is a different story. Here is how it compares to Bordair.

What PromptGuard is

PromptGuard is a fine-tuned mDeBERTa model released by Meta under the MIT licence. It classifies text inputs as benign, injection, or jailbreak. It is available on Hugging Face and can be self-hosted.

The self-hosting trade-off

PromptGuard gives you the model weights, but you are responsible for everything else: infrastructure, scaling, monitoring, updating, and extending to new modalities. For a small team shipping an LLM feature, this is significant overhead.

Bordair is a managed API. One HTTP call, sub-50ms response, no infrastructure to maintain. We handle model updates, scaling, and monitoring so you can focus on your product.

Multimodal detection

PromptGuard is text-only. If your application accepts images, documents, or audio (and most modern ones do), you need to build your own extraction and detection pipeline for those modalities.

Bordair handles text, image, document, and audio natively. Our pipeline extracts content from every modality and runs it through the classifier. Cross-modal attacks, where the injection is split across channels, are detected automatically.

Detection quality

PromptGuard was trained on a broad dataset but has known issues with false positives on certain input patterns. Bordair's three-stage pipeline (pattern match, fast-accept, ML) reduces false positives by catching obviously benign inputs early.

Output scanning

PromptGuard does not include output scanning. Bordair offers regex-based output rules on paid plans, letting you block, redact, or flag sensitive content in LLM responses.

When to choose PromptGuard

• You want full control over the model and infrastructure
• You have ML engineering resources to maintain and improve the classifier
• You only need text-based detection
• You want to avoid any external API dependency

When to choose Bordair

• You want a managed solution with no infrastructure overhead
• You need multimodal detection
• You want output scanning
• You need production-grade latency and reliability
• You want to ship prompt injection protection in minutes, not weeks

Try Bordair free with 1,000 scans per month.