The UK Institute for AI Security is working with international partners to lead a £15 million project focused on investigating AI coordination.
The alignment project also features the Canadian Institute of AI Safety, the Canadian Institute of Advanced Research (CIFAR), Schmidt Science, Amazon Web Services, Humanity, Halcyon Futures, Safe AI Funds, UK Research Innovation, Advanced Research and Invention Agency (ARIA).
They are the pioneer of new works designed to always function as intended by AI systems. This is an area that is becoming increasingly important as AI systems become more sophisticated and autonomous.
Inconsistency broadly refers to AI systems that act against developer goals, policies and requirements. That's intentional. That is, a threat actor who destroys AI systems to attack a target – or if an inconsistency occurs due to the inability to implement the proper AI guardrail.
Read more about AI Safety: OWASP launches agent AI security guidance
According to Trend Micro, examples of inconsistencies include:
- Model addiction: Attackers inject or manipulate LLM training data, leading to biased output, false decisions, and sometimes backdoors
- Rapid Injection: Threat Actors create malicious prompts that overcome LLM's built-in guardrails, affecting system jailbreak
- Contingent Disclosure: Inadequately designed AI systems may incorrectly access and share privileged information with users
- Runaway Resource Consumption: If resource consumption is not properly bounded, AI components may tackle sub-problems in a self-replica manner and administer the system
Peter Kyle, secretary of science, technology and innovation, said advanced AI systems are already beyond humanity in some regions, making the project more urgent than ever.
“All AI integrity is being acted in our best interest as the system is prepared to make sure we work the way we want it to. This is at the heart of the work our lab has been leading from day one, protecting national security and ensuring that the British people are protected from more serious risks.”
“Responsible AI development requires a coordinated, global approach. This fund will help AI to provide more reliable, more reliable, and grow, better public services and highly skilled jobs.”
