Task-force-for-AI-agents-in-Healthcare/haarf

The framework operationalizes 279 requirements across 8 verification categories through a modular Python evaluation harness with multi-layer middleware enforcement (RBAC, contraindication checking, injection sanitization, circuit breakers, audit logging) that runs red-team scenarios against LLM agents from Anthropic and Google. It synthesizes FDA, EU AI Act, NIST, WHO, and ISO/IEC standards into a provider-agnostic tool-use agent loop, with experimental validation showing deterministic elimination of unauthorized tool execution (56–60% baseline → 0% under HAARF) across 600 trials and cross-model validation.