International Collaboration: Surrey, Sheffield, and Melbourne Join Forces on £3.2m Project to Create Provably Secure Sandbox for AI Agents

AI Security: Surrey-Led Project to Develop Secure Digital Sandbox for AI Agents

A groundbreaking initiative led by Surrey University aims to design a digital sandbox that will provide a secure and tightly controlled environment for AI agents. This project has recently secured over £3.2 million in funding from the Advanced Research + Invention Agency (ARIA) under its Safeguarded AI programme.

The need for better safeguards for AI agents has become increasingly evident, especially in light of recent incidents involving Anthropic and OpenAI. The Agentic-seL4 project, spearheaded by Professor Brijesh Dongol from Surrey’s Centre for Cyber Security, will delve into the containment of AI agents to ensure their safe operation.

The project aims to create a rigorously verified sandbox that regulates the access and actions of AI agents, allowing for dynamic permissions that can be adjusted or revoked in real-time. This will enable organizations to control and monitor AI agents more effectively, ensuring that they operate within predefined boundaries.

Formal models and proofs will underpin the sandbox system, guaranteeing that specified policies are consistently enforced to prevent data leakage. Additionally, a proof-of-concept reference design will be developed and tested against various scenarios to validate the effectiveness of the restrictions imposed.

Empowering AI Agents within Safe Boundaries

Professor Brijesh Dongol emphasized the importance of adapting safeguards to keep pace with the rapid advancements in AI technology. The project’s goal is to grant AI agents the freedom to perform tasks efficiently while maintaining trust in their operations.

The collaborative effort between Surrey, Melbourne, and Sheffield Universities marks a significant milestone in AI research. The Agentic-seL4 project is part of ARIA’s Safeguarded AI programme, which aims to leverage AI-enabled formal methods for enhanced cybersecurity defense.

By exploring the intersection of AI and cybersecurity, researchers hope to make high-assurance cyber defense more accessible and practical than ever before. The ultimate objective is to establish a robust framework that safeguards critical systems against potential threats posed by AI technology.