The world “has no time” to prepare for the security risks posed by cutting-edge AI systems, according to a leading figure at the UK government’s scientific research agency.
David Dalrymple, the area agency’s program director and AI security expert, told the Guardian people should be concerned about the growing potential of the technology.
He said, “I think we should be concerned about systems that can do all the work that humans do to function in the world, but in a better way.” “We will fall behind in all the areas in which we need to be effective in order to maintain control over our civilization, society, and planet.”
Dalrymple said there is a gap in understanding between the public sector and AI companies about the power of imminent breakthroughs in the technology.
“I would advise that things are moving really quickly and from a security perspective we may not have time to get ahead of this,” he said. “And it is not science fiction to project that within five years most economically valuable tasks will be performed by machines at a higher level of quality and lower cost than humans.”
Dalrymple said that governments should not assume that advanced systems are reliable. ARIA is publicly funded but independent from the government and directs research funding. Dalrymple is developing systems to secure the use of AI in critical infrastructure such as energy networks.
“We cannot assume that these systems are reliable. The science to do so is unlikely to materialize in time given the economic pressures. So the next best thing we can do, that we may be able to do in time, is to control and mitigate the downside.”
Describing the consequences of technological advancements that outpace security as “destabilization of security and the economy”, Dalrymple said that more technical work is needed to understand and control the behavior of advanced AI systems.
“Progress can be considered destabilizing and it can actually be good, at the limit of what many people are expecting. I’m trying to make things better but it’s too big a risk and human civilization as a whole is going through this transition.”
This month the UK government’s AI Security Institute (AISI) said the capabilities of advanced AI models are “rapidly improving” across all domains, with performance in some areas doubling every eight months.
According to the institute, leading models can now complete trainee-level tasks an average of 50% of the time, up from about 10% last year. AISI also found that the most advanced systems can autonomously complete tasks that would take a human expert more than an hour.
The institute also tested advanced models for self-replication, which is a major security concern because it involves a system that spreads copies of itself to other devices and becomes harder to control. The tests showed that the two state-of-the-art models achieved a success rate of more than 60%.
However, AISI stressed that worst-case scenarios are unlikely in an everyday environment, and said that any attempts at self-replication “are unlikely to succeed in real-world conditions”.
Dalrymple believes that AI systems will be able to automate the equivalent of an entire day of research and development work by the end of 2026, resulting in “further acceleration of capabilities”, as the technology will be able to self-improve on the mathematics and computer science elements of AI development.
