Anthropic 称反蒸馏封号已阻断 99% 窃取尝试
转述 Anthropic 说法：代理网络操控数十万假账号、有实验室发 1.2 万次套取推理请求，数万账号被封且常在 2 小时内，误判率低于 0.1%。
Anthropic says it has cut attempts to steal claude's intelligence by up to 99% and there's a solid reason why Claude took this approach the aim goal is to stop rival labs from using millions of prompts to copy Claude's capabilities, especially its reasoning here's what's interesting from the look of things: > proxy networks running hundreds of thousands of fake accounts > one lab reportedly sent 12,000 requests designed to extract Claude's reasoning > tens of thousands of accounts banned, often within 2 hours > less than 0.1% of legitimate requests wrongly flagged, according to anthropic therefore, Anthropic stated that it is blocking trillions of tokens in paid traffic to protect its models what are your thoughts on this guys?