Researchers propose a formal security model to ensure large language models safely use external tools without unauthorized access or interference between user requests.

Introduces a formal authorization framework for controlling how LLMs access and execute tool schemas

Provides noninterference guarantees to prevent one user's tool usage from compromising another user's security or data

Addresses growing security concerns as LLMs become integrated with external APIs and applications

Combines formal verification methods with practical constraints for real-world LLM deployment scenarios