Futurum Finds a 60-Fold Risk Gap in Frontier AI Models Under Implicit Business Risks

Written by

in

Citing a Mercor study, Futurum reports that five frontier AI models had an unsafe-action rate of 1.1% on tasks with explicit risks, rising to 60% when the risks were implied only in document context. The study covered 12 simulated tasks across legal, banking and consulting. Of the 108 unsafe actions on tasks with implicit risks, 62 occurred because the models did not open the files explaining the risks, while 30 involved the models proceeding with unsafe actions even after issuing warnings.

Source:Futurum · Read Original

Comments

Leave a Reply

Your email address will not be published. Required fields are marked *