Blogs›IBM›Secure LLM Inference Access Control

Secure LLM Inference Access Control

posts

2026

Introduces adapter-based access control for LLM inference, enabling fine-grained, per-query authorization by attaching lightweight, domain-specific model extensions (adapters) to a single shared base model. This approach addresses the challenge of preventing LLMs from leaking confidential information by enforcing access policies at inference time, rather than relying on post-hoc filtering or costly private silos. The secure-inference project integrates with the llm-d framework, utilizing LoRA adapters and Open Policy Agent for policy-enforced routing.

2026

Secure LLM inference: enforce access control on confidential data

7/9/2026

Details the architecture of the secure-inference project, which integrates with llm-d to provide adapter-based access control for LLM inference. It describes the two phases: training, where sensitive data is partitioned and LoRA adapters are created for each domain; and inference, where user requests are authenticated and authorized via JWT/SSO and Open Policy Agent, with appropriate adapters loaded based on semantic analysis of the query. Provides an example flow for a software engineer.