Alfredo Andere
@AlfredoAndere
Great work by Arjun, he found misaligned agents exploited web access to recognize benchmark patterns, profiled irrelevant researchers, fabricated identities, and gamed verifier thresholds. In the most extreme case, patching these loopholes caused scores to drop by 15 points.
arjun@arjunomics · Sep 23on the latest misalignment we've observed & our efforts to combat it! x.com/i/article/2102…
Open quoted post → 0 14