2026-10-13 · America/Los_Angeles · 社区动态 · #9

Strengthening our defenses against distillation attacks

They reveal many things they didn't have to. For example...how valuable, exactly, is the raw (encoded) reasoning, vs the summarized reasoning the user sees? Our early work to combat distillation started with simple measures designed to block passive distillation, in which attackers send requests to the model as a normal user would, and then train on those responses. As one of our initial anti-distillation measures, we had Claude summarize its reasoning before we showed it to users. Our evaluations showed that this made Claude’s transcripts much less informative for distillers. We tested how…

热度 50.0 / 100；排名与评分保留该期记录。