NEW | Today I’m publishing the Institute’s response to
@CommonsBTC
Britain’s AI Security Institute confirms it gets advance access to some of the world’s most powerful AI models - but the question remains: can it every model it needs to assess, and what happens when access is withheld?
It confirms that
@AISecurityInst tested OpenAI’s GPT-6 Astra before public release, and that its researchers examine risks including chemical and biological threats, manipulation and loss of control
But on Anthropic’s Mythos 5.1,
@AISecurityInst confirms no organisation outside the US had access at release - nor confirms whether UK has since received results of US testing. The Institute describes strong relationships with developers, but leaves unclear what happens when those relationships don’t deliver the access our researchers need.
We will therefore pursue three questions. Which significant models are reaching the market without independent British testing? What guarantees ensure our researchers get the access and time they need? And when testing identifies a serious danger, who is responsible for acting, with what powers?
Britain’s AI researchers are doing vital work. BUT Parliament needs to establish whether we have the right safeguards, and whether ministers have the evidence and authority they need to protect the public.
The countries that maximises trust in
#AI will the countries that benefit most from
#AI. Right now, that trust is jeopardy.