The world “may not have time” to prepare for the safety risks posed by cutting-edge AI systems, according to a leading figure at the UK government’s scientific research agency.

David Dalrymple, a programme director and AI safety expert at the Aria agency, told the Guardian people should be concerned about the growing capability of the technology.

“I think we should be concerned about systems that can perform all of the functions that humans perform to get things done in the world, but better,” he said. “We will be outcompeted in all of the domains that we need to be dominant in, in order to maintain control of our civilisation, society and planet.”

Dalrymple said there was a gap in understanding between the public sector and AI companies about the power of looming breakthroughs in the technology.

“I would advise that things are moving really fast and we may not have time to get ahead of it from a safety perspective,” he said. “And it’s 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 by humans.”

Dalrymple said 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 safeguard AI’s use in critical infrastructure such as energy networks.

“We can’t assume these systems are reliable. The science to do that is just not likely to materialise in time given the economic pressure. So the next best thing that we can do, which we may be able to do in time, is to control and mitigate the downsides,” he said.

Describing the consequences of technological progress getting ahead of safety as a “destabilisation of security and economy”, Dalrymple said more technical work was needed on understanding and controlling the behaviours of advanced AI systems.

“Progress can be framed as destabilising and it could actually be good, which is what a lot of people at the frontier are hoping. I am working to try to make things go better but it’s very high risk and human civilisation is on the whole sleep walking into this transition.”

This month the UK government’s AI Security Institute (AISI) said the capabilities of advanced AI models were “improving rapidly” across all domains and the performance in some areas was doubling every eight months.

Leading models can now complete apprentice-level tasks 50% of the time on average, up from approximately 10% of the time last year, according to the institute. AISI also found that the most advanced systems can autonomously complete tasks that would take a human expert over an hour.

The institute also tested advanced models for self-replication, a key safety concern because it involves a system spreading copies of itself to other devices and becoming harder to control. The tests showed two cutting-edge models achieving success rates of more than 60%.

However, AISI stressed that a worst-case scenario was unlikely in a day-to-day environment, saying any attempt at self-replication was “unlikely to succeed in real-world conditions”.

Dalrymple believes that AI systems will be able to automate the equivalent of a full day of research and development work by late 2026, which will “result in a further acceleration of capabilities”, because the technology will be able to self-improve on the maths and computer science elements of AI development.