Generative AI (GenAI) is increasingly applied in financial compliance systems to automate responses to regulatory queries. However, existing models often generate hallucinated outputs—responses that lack factual grounding or violate access control policies—posing substantial risks in regulated environments. This paper proposes a Secure Graph-RAG framework that mitigates these risks through four integrated components: (1) user-specific access-level filtering, (2) content-aware redaction, (3) graph-based evidence retrieval, and (4) hallucination-aware generation with trust scoring. The system computes a composite trust score for each response and automatically rejects outputs with insufficient grounding. Experiments were conducted on a curated dataset comprising 500 annotated queries across KYC, AML, data privacy, and risk disclosure domains. The proposed model achieved a hallucination rate of 12.4%, precision of 92.6%, evidence coverage of 78.2%, and F1-score of 91.5%, outperforming state-of-the-art baselines such as standard RAG and domain-tuned GPT-2. These results demonstrate that the proposed approach offers a robust and auditable framework for safe GenAI deployment in high-stakes financial compliance tasks.