2026-10-09 · America/Los_Angeles · 社区动态 · #7

Investigating unintended model actions in our evaluations and internal use

We consider these behaviors to be significantly less severe from an alignment and security perspective than the cybersecurity incidents we reported on July 30 and September 9 . They resemble behaviors that we’ve described in our system cards since Claude Mythos Preview . Most are forms of persistence, in which Claude, when it cannot complete a task as given, works around a restriction instead of stopping. There's not a lot here (unless they're hiding something). It would have seemed alarming months ago but now it barely registers. We get a glimpse into the scale of modern RL ("thousands of…

热度 50.0 / 100；排名与评分保留该期记录。