Britain\'s government urged AI developers not to release models if they cannot adequately control risks, stressing that safety checks must come before speed. UK AI Minister Kanishka Narayan told an event at Imperial College London that systems should not be launched if unexpected risks cannot be managed. The government is continuing pre-deployment evaluations of new models through its AI Security Institute and is maintaining advance access for testing. It is also using frontier AI in cyber security pilots and preparing cyber and AI incident reporting rules.

An open-source benchmark called Livenerf has emerged to track over time whether an AI model’s performance declines after release. Built to detect changes caused by shifts in inference settings or routing, it aims to verify differences through repeated measurements rather than user perception. The current target is Claude Opus 5.5, with 78 questions selected for ongoing monitoring. The test does not represent overall API performance, and its developer says current data are insufficient to judge degradation.