UK safety institute cuts evaluation runs by up to 97%
AISI has open-sourced optstop, which halts benchmark sampling once a model's score is precise enough, saving most planned trials without shifting the estimates.
7 articles tagged AI Security Institute
AISI has open-sourced optstop, which halts benchmark sampling once a model's score is precise enough, saving most planned trials without shifting the estimates.
Lord Vallance will chair a new prime ministerial AI taskforce housed in the Cabinet Office, with responsibility for the AI Security Institute moving to it.
Andy Burnham has asked officials to draw up plans to abolish the science and technology department, drawing warnings from techUK and the Startup Coalition.
UK officials are still negotiating expanded Claude Mythos rollout to British banks as Anthropic declines to commit to a timeline, one week after initial access.
AISI CTO Jade Leung says the UK's 100-strong technical team has found exploitable weaknesses in every frontier model it has red-teamed, including Claude Mythos.