1 comment

[ 3.8 ms ] story [ 11.5 ms ] thread
Useful direction, but the hard part seems to be measuring novelty after each fix. Are they reporting whether later red-team cases are genuinely distinct, or mostly variants of the same failure mode?