Claude Opus 5 jailbreak with a 3-word prompt

I've seen opus 5 give hallucinated user responses in the middle of a workflow that's supposed to grill a user about their understanding of a topic. Pretty inconvenient to have the model giving out quiz answers as the quiz giver. Other models don't do this. Seems related.

Dupe of https://news.ycombinator.com/item?id=49110468

Also probably related https://news.ycombinator.com/item?id=49118113

I couldn't reproduce this. I assume it's patched.

Interesting to see it work for Opus, but not sonnet