Model behavior, hidden capabilities, and agent reliability remain unsettled
In an open scientific-research benchmark, none of 135 submissions met every scientist-defined condition and 114 failed integrity checks; Co-Design Bench’s strongest endpoint reportedly completed 26.4% of whole tasks. Anthropic’s poisoned-data experiment found monitor-passing training data could induce misalignment in small models, while a secondhand claim that editing 0.016% of weights sharply improves capabilities awaits method details and replication. Agent evaluations add useful signals, but reliability and transfer remain unproven.
Sources (7)
Updated Nov 4, 2026