What happened
Around 23 September 2026, CETaS (Centre for Emerging Technology and Security, part of the UK's Alan Turing Institute) published a paper by Prof. Rick Hennessy and Dr. Carolyn Ashurst proposing a sociotechnical framework for the behavioural assurance of agentic AI. It introduces a taxonomy of behavioural failure modes spanning individual agents, multi-agent systems and human-agent interaction, identifies access/authority/trust as the three factors that turn agent behaviour into organisational harm, and sets out seven behavioural assurance principles. A detailed operational report was flagged as forthcoming.
Why it matters
It is the first publication in a UK national-institute project that gives sensitive/high-stakes organisations (government, critical sectors) a vocabulary for assuring agent behaviour beyond model evaluation — complementing the 'agents act but developers detect late' incident pattern and feeding into UK assurance and governance frameworks.
Action needed
CISOs and AI-governance teams in high-stakes UK/regulated settings should use the failure-mode taxonomy and access/authority/trust lens when scoping agentic AI oversight and intervention controls.