The long-expected news came last week. The FT reported that the UK’s AI Security Institute (AISI) had not been offered access to Anthropic’s latest large language model (LLM), Claude Mythos 5.1. For the institute, widely lauded as a successful UK initiative and described internally as “a start-up in the government”, this was a humbling moment.
But Anthropic’s exclusion of AISI from testing access shouldn’t necessarily surprise us. Before that, the institute had been offered unique access to pre-deployment models through a system of patronage dependent on the good grace of both the American frontier companies and now, officials suspect, the US government. Could Britain’s new normal now involve no access to cutting-edge LLMs?
In the same week, Anthropic researcher Jacob Coxon resigned, writing that he was deeply concerned about AI technology wiping out humanity. Evan Hubinger, a fellow Anthropic employee, responded by saying that he agreed with Coxon’s concerns and that there was a 10% chance of AI killing all humans.
There is legitimate concern among both the public and the British political class about these developments. How could there not be? But concerns surrounding AI safety are being misdirected in ways likely to make us more vulnerable.