UK AI Safety Institute Releases Evaluation Report: GPT-4 and Claude Perform Well in Bioweapon Knowledge Tests

The UK AI Safety Institute (AISI) released a safety evaluation report on several advanced AI models, testing their behavior in high-risk scenarios such as bioweapon creation and cyberattack assistance. The report shows that Claude and GPT-4 performed well in safety tests, but some emerging open-source models have significant safety gaps. The report advances the standardization of international AI safety assessments.