TL;DR
The Alan Turing Institute has published a report, “Frontier AI Risks: A practical way forward”, which says disputes over humanity’s long-term fate should not hold up work on dangers that can already be measured. It picks out five risk areas and says the UK’s next task is checking whole AI systems after deployment, not just models before release. A £2 million philanthropic grant will fund a linked research programme.
Five risks, one method
The report’s priorities are chemical and biological misuse, cyber attacks, misalignment, threats to democratic stability, and the loss of meaningful human control. Its case is that many of the capabilities and harms behind each can be seen and studied today, so they can be acted on without first settling the existential debate.
The central proposal is about verification. Britain has built a strong record in evaluating frontier models ahead of launch, the Turing says, but a model tested on its own tells you little about how a full system behaves once people, processes and changing conditions are wrapped around it. It wants more work on assuring deployed systems in real settings, and says it will work with the NCSC and the AI Security Institute on it.
Two popular ideas get a cooler reception. Kill switches would only work with governance, fallback arrangements and operating procedures behind them, and a real slowdown in frontier development would need a level of cooperation between regulators, industry and governments that the report judges improbable today.
“The risks posed by advanced AI are real, but they are not beyond our control,” said Professor Jason McEwen, the institute’s interim chief scientist.
Money and a second paper
CETaS, the institute’s security research centre, has won £2 million from Coefficient Giving for research on how the most powerful AI could change the security picture and how government can prepare. CETaS also released a briefing on agentic AI, listing eight ways agents can act against an organisation’s intentions and seven principles for governing them.
Looking forward
The report lands in a crowded week. Ed Davey wants a treaty to stop super-intelligent AI, 22 countries have backed a declaration on human control that the UK has not signed, and AISI testing remains voluntary, as shown when Anthropic released a model without it. The Turing’s pitch sits between those positions: less about pausing, more about proving deployed systems can be trusted. Whether ministers fund that kind of assurance at national scale is the open question.