The Government’s AI Safety Institute is to open an overseas office in San Francisco to get closer to the tech giants based in Silicon Valley. The Department for Science, Innovation and Technology (DSIT) said the new office is expected to open in the summer, and would look to engage with the world’s biggest AI labs, which are based in the region. The UK’s AI Safety Institute was first announced at the AI Safety Summit at Bletchley Park last year and, alongside several similar international institutes, will carry out research and conduct evaluations of AI models to help inform AI safety policy. The institute has released its first selection of recent results from safety testing five publicly available advanced AI models, although the institute said it had anonymised the results. In its findings, the institute said several models had completed cyber security challenges but struggled to complete more advanced challenges, and several models had also shown similar to PhD-level knowledge of chemistry and biology. The institute said all the models it tested remain highly vulnerable to basic “jailbreaks” to sidestep safety controls, and some will produce harmful outputs even without dedicated attempts to get around safety mechanisms. The testing also noted...