NeuralStackThe UK Department for Science, Innovation and Technology released AI safety testing regulations on July 14. The rules apply to any model trained with more than 10^26 floating-point operations. Mandatory third-party evaluations begin October 1 2026.
Approved testing labs must assess dangerous capability benchmarks including biological weapons planning. Companies have 45 days after training completion to submit results. Fines reach 6 percent of global annual revenue for violations.
The policy follows the 2025 AI Safety Summit commitments. Similar frameworks exist in the EU and proposed US legislation. The UK aims to become a global leader in responsible AI governance.
Why this matters
The mandate creates the first binding pre-deployment testing requirement outside the EU. Frontier labs must allocate additional resources for compliance documentation. This raises the cost of releasing new large models.
Smaller developers gain clarity on regulatory thresholds. International alignment efforts may accelerate through shared testing standards. The policy could influence upcoming US and Chinese regulations.
Implementation guidance will be published by August 20. The government plans to expand covered capabilities in 2027.