UK AI Security Institute: OpenAI and Anthropic models raised serious concerns in testing

The UK's AI Security Institute stated that flagship models from OpenAI and Anthropic raised "serious concerns" during state testing. Details are sparse as usual, but such wording from a body with pre-release access to the models outweighs a hundred blogger reviews.

The trend matters: independent state examination of AI is institutionalizing — the UK and Singapore already run such centers, and the EU is building its own under the AI Act. The AI market is taking on pharmaceutical traits: the maker must prove safety to an outside regulator, not just describe it in press releases. The gap between vendor marketing and testers' conclusions is becoming public in the process.

For a business in Kazakhstan choosing models for its tasks, the practice follows: when selecting an AI platform, look beyond price and speed to independent evaluation results — they are published and updated. And for critical processes, design in the ability to swap models: a vendor whose model fails the next audit must not drag your processes down with it. Architectural portability has stopped being engineers' over-caution and become business insurance.