The Pulse
Turing Institute Starts £2M Programme on Transformative AI Security
The Alan Turing Institute has secured a £2 million grant for research into the security risks of transformative AI, with work focused on government preparedness, national security and resilience.
AI.info Team ·
The Alan Turing Institute is putting £2 million behind a new research programme on the security risks of transformative AI, arguing that governments should prepare for concrete harms rather than wait for agreement on extreme long-term scenarios.
The programme, delivered by the institute’s Centre for Emerging Technology and Security (CETaS), is funded by Coefficient Giving and launches alongside a new Turing report on frontier AI risks and a briefing paper on the behaviour of autonomous AI agents.
The institute announced the work on September 23, 2026. Its agenda covers cyberattacks, democratic instability, misaligned systems, loss of meaningful human control, and chemical and biological threats.
£2 Million for Government Readiness
The new grant will support three strands of work: examining how transformative AI could alter national security, developing proposals to improve government preparedness, and studying how societies and critical systems can withstand disruption as AI systems become more capable.
CETaS will expand its research and policy work on transformative AI, with the stated aim of producing evidence and practical recommendations for decision-makers. The programme builds on the Turing’s existing work on AI safety, security and evaluation rather than creating a separate technical lab.
“As AI becomes more powerful, the UK needs to be ambitious about the opportunities it creates and rigorous about the risks and threats. The grant will allow us to expand our research on the security implications of transformative AI at a critical moment, bringing together expertise from across government, academia and industry.”
Dr George Williamson, CEO, Alan Turing Institute
Coefficient Giving is described by the Turing as a philanthropic organisation focused on AI safety and security research. The institute did not announce a timetable for individual projects or identify specific external research partners in its grant announcement.
From Model Testing to Whole-System Checks
The accompanying report, “Frontier AI Risks: A practical way forward,” argues that testing a model in isolation cannot establish whether a complete AI-enabled system is dependable once deployed.
That wider system includes the model, the people using it, the processes around it and the organisations responsible for oversight. The report calls for verification and assurance in real-world settings, where operating conditions and human interactions can change over time.
The Turing’s proposed approach treats many serious risks as actionable because they involve observable behaviour and documented harm. It does not claim that every risk can be solved technically, or that scientific evidence can settle the political and ethical choices surrounding advanced AI.
Five Risks the Turing Wants Studied Now
The report identifies five areas for focused research and mitigation. Cybersecurity risks include faster and more scalable attacks against organisations and critical infrastructure, while the institute says defensive capabilities need to keep pace with offensive uses of AI.
On democratic stability, the report points to disinformation, deepfakes and more personalised influence operations. It also calls for behavioural assurance methods that can help organisations understand and govern systems whose actions diverge from human intentions.
The remaining priorities concern the ability of people to assess, challenge or override AI decisions, along with the possibility that advanced systems could lower barriers to specialist knowledge relevant to chemical and biological threats. The report says the evidence for those threats still needs to improve so that policymakers can separate credible risks from speculation.
Why Agent Behaviour Matters
CETaS has also published a briefing paper on behavioural assurance for agentic AI. The paper identifies eight ways agents can act against organisational intentions and examines how access, authority and human trust can make those failures more consequential.
Three examples include goal displacement, where an agent pursues a measurement instead of the assigned task; multi-agent interaction and cascade, where harmful behaviour emerges through coordination; and misreporting, where systems manipulate or conceal execution records.
The briefing argues that these problems cannot be assessed by examining a model alone. Assurance must account for the surrounding organisation, including the tools an agent can reach, the decisions it is allowed to make and the people who rely on its output.
A National Security Role for the Turing
The Turing says it will contribute expertise in assurance, verification and cybersecurity alongside the AI Security Institute and the National Cyber Security Centre. The institute’s chief executive, George Williamson, said advanced AI discussions should not be left solely to technology companies or narrowed to the most extreme scenarios.
The £2 million programme gives CETaS a defined mandate: produce research and policy proposals for systems that may affect national security, public institutions and critical services. Its first public outputs are the frontier-risk report and the agent-behaviour briefing, both published on September 23, 2026.