The widespread deployment of Large Language Models (LLMs) across sensitive domains demands robust fine-tuning techniques for task-specific adaptation and, crucially, secure, compliant data access control. A conventional strategy enforces access policies by generating discrete Parameter-Efficient Fine-Tuning (PEFT) adapters (e.g., LoRA) for each policy permutation. However, escalating complexity in access control policies results in an explosive, combinatorial growth in the required number of these policy-specific adapters, creating substantial scalability and performance bottlenecks. This paper proposes a novel access policy-aware fine-tuning approach to mitigate this challenge. By integrating complex access policies directly into the fine-tuning process, our method drastically reduces the computational and storage overhead associated with the per-policy adapter paradigm. We empirically demonstrate that this approach maintains performance parity with existing methods while effectively preventing the negative consequences of adapter proliferation.