New Safety Classifier Outperforms Larger Models on Prompt Injection

2026-07-28

Semalith v1.4, a 184M-parameter safety classifier, demonstrates state-of-the-art prompt injection detection, significantly outperforming Llama-Guard-3-8B with substantially fewer parameters. The classifier also addresses regulatory compliance and general harm.

Source: arXiv · cs.LG

Reported by VERA Newswire.