Beyond Injection Detection: A Positive-Security Prompt Firewall that Closes the Scope and PHI Gap SOTA Classifiers Miss in Healthcare
This paper introduces QFIRE, a high-performance Rust-based prompt firewall that addresses critical gaps in healthcare AI safety by combining positive security scope constraints, PHI-specific detection, and de-obfuscation techniques to effectively block legitimate-looking data exfiltration and out-of-scope requests that state-of-the-art injection classifiers miss, all while maintaining low latency and deterministic auditability.