比起所有这些评估分数,对我来说最令人兴奋的是另一件事:Opus 5 是我们迄今为止最难以进行提示注入的模型。它在系统卡中有点隐藏,但在 PI 评估和红队测试中,成功进行提示注入在 Opus 5 上非常困难。
— Boris Cherny:https://twitter.com/bcherny/status/2080713091688583312,这里是系统卡的相关部分:https://www-cdn.anthropic.com/c5fbac3f0b1280a933ebd26d3cb8bb9f5bdeaf48/Claude%20Opus%205%20System%20Card.pdf#page=73,第73页
这是 Simon Willison 收集的引用,于2026年7月25日发布:/2026/Jul/25/。
More than any of these eval scores, what is most exciting to me is something else: Opus 5 is our least prompt injectable model yet. It is a bit buried in the system card, but across PI evals and red teaming, Opus 5 is very hard to prompt inject successfully.
— Boris Cherny:https://twitter.com/bcherny/status/2080713091688583312, here's that System Card section:https://www-cdn.anthropic.com/c5fbac3f0b1280a933ebd26d3cb8bb9f5bdeaf48/Claude%20Opus%205%20System%20Card.pdf#page=73, page 73
This is a quotation collected by Simon Willison, posted on 25th July 2026:/2026/Jul/25/.