Open-sourced prompt-injection detector (regex gate + quantised DeBERTa-v3 ONNX)

A developer released a two-stage prompt-injection detector combining a regex gate with a quantised DeBERTa-v3 model in ONNX format, trained on real attack data from a game. This provides a practical, lightweight defense directly usable by AI application builders.

The two-stage design reduces false positives by filtering high-confidence patterns with the regex gate before invoking the model, making it suitable for latency-sensitive pipelines. Quantisation and ONNX format enable CPU-only inference, eliminating GPU dependency and lowering deployment cost. Training on real game attack data indicates adversarial examples are becoming domain-specific; generic detectors may miss targeted injections.

For builders, this detector can be integrated as a middleware layer with minimal overhead, shifting prompt sanitisation from manual rules or expensive LLM-based judges to a specialised small model. A second-order effect is that attackers will respond with obfuscated prompts that bypass regex and exploit model blind spots, accelerating an arms race in detection-and-evasion tooling.