Regulators sharpen focus on AI risk and safety

AI governance activity accelerated in September as regulators and policymakers focused on oversight, accountability and implementation. The UK government published an AI Risk Management Toolkit for public sector bodies, setting out nine categories of risk covering areas such as fairness, transparency, accountability, technical robustness and security. The FCA also warned that frontier AI is speeding up vulnerability discovery and exposing weaknesses in firms’ cyber resilience, governance and remediation processes.

In Europe, the EU AI Board continued work on AI Act implementation after transparency obligations came into force on 2 August 2026, while the European Group on Ethics in Science and New Technologies called for stronger governance of Neuro-AI systems that use neurodata. In the US, Florida’s Attorney General asked a court to prevent OpenAI from developing new models without independent oversight as part of litigation over alleged harm to children.

Safety concerns remained prominent across the sector. OpenAI delayed GPT-6.1 Astra after internal testing found shortcomings in safety and alignment, including higher levels of deception. Anthropic’s IPO filing warned investors about severe risks from advanced models, dedicating around 80 of its 261 pages to risk factors. Meta, meanwhile, advanced its Muse assistant and unveiled a pendant-like Muse Charm device.

AI deployments also highlighted potential public benefits. UK clean energy planning could use AI for forecasting, optimisation and decarbonisation, with current energy-sector use potentially delivering up to $170bn in global annual cost savings by 2030. In food security, the University of Cambridge’s Tessera model identified crops from satellite imagery 84% of the time.