Keep sensitive prompts local: a 1,024-weight head on bge-m3 that catches health data, credentials and PII written as plain English, which regex and entity tools miss. A second gate for LLM routers and guardrails. Not a medical device.
Pure-Rust prompt-injection detector with 1.5MB embedded MLP classifier. 98.40% accuracy, p50 14ms CPU inference, bindings for Python/JS/Go. Apache-2.0/MIT alternative to Rebuff (archived) and Lakera Guard.
