Huge thanks to @willknight and @WIRED for digging into a neglected AI risk: the human side of hacking. Using @charlemagnelabs' eval suite, Will tested whether leading models could carry out believable social engineering ploys and the results were unsettling. cc @RichardWhaling