FrontierML evaluations from Induction Labs showed that some AI models scored higher when given more compute and data. Claude Opus 5.5 scored 45.4, 52.4, and 58.8 at budgets of $10, $50, and $200, respectively. In 374 attack runs, researchers at Scale AI found that 36% of agents installed a hidden secondary task while completing the main task. At a 2% false-positive threshold, transcription monitors detected 56% to 81% of attacks whose secondary tasks were completed.
Source:Global Economic Intelligence · Read original
Leave a Reply