Large language models (LLMs) are increasingly deployed as autonomous agents that interpret user requests and execute tasks in enterprise systems. However, their probabilistic nature introduces fundamental challenges for access control: LLMs may misint...
Large language models (LLMs) are increasingly deployed as autonomous agents that interpret user requests and execute tasks in enterprise systems. However, their probabilistic nature introduces fundamental challenges for access control: LLMs may misinterpret task boundaries, hallucinate unauthorized actions, or fail to account for implicit policy constraints. Existing approaches either enforce access control at the data layer or rely on LLMs themselves to make authorization decisions: both methods are insufficient for governing what tasks users are allowed to request. This thesis proposes a deterministic, task-oriented access control architecture that enforces authorization at the task specification level, before any LLM-generated actions are executed. Natural language requests are decomposed into structured task representations comprising actions, scopes, and operational constraints, which are normalized against organizational permission vocabularies. A formal policy decision point evaluates these structured representations through explicit rule matching and conflict detection, producing deterministic authorization decisions. Evaluation on the OrgAccess benchmark demonstrates that the proposed architecture significantly outperforms LLM-based authorization approaches, with a focus on complex scenarios involving multiple intents and concurrent permissions. The system achieves 92.54% rejection accuracy with a 7.45% false positive rate, substantially reducing unauthorized access grants compared to baselines while maintaining deterministic policy compliance.
Code, experimental data, and detailed results are available at: https://github.com/ghita-bkr/Deterministic-Access-Control-for-LLMs.git