UK AI Safety Institute Warns Frontier AI Risks Are Shifting to Development Stage

The UK Artificial Intelligence Safety Institute published an article on October 12 stating that as frontier AI capabilities advance rapidly, safety risks are moving earlier into the model development phase, and existing safety evaluation methods are struggling to keep pace. It called for exploring new technical safety practices to ensure frontier AI development proceeds safely.

The article noted that previous safety assessments mainly focused on ensuring models were safe at the time of public release. However, some AI systems now show increased autonomy, particularly in the cybersecurity domain, and may pose risks during training, internal deployment, and evaluation even before they are publicly released.