Claude 5 family's hallucinations look a lot like internal Anthropic emails

Posted by arm32 15 hours ago

Counter3Comment4OpenOriginal

Comments

Comment by arm32 15 hours ago

This is the full dataset related to the blog post at: https://alec.is/posts/exploring-the-dario-and-amanda-prompt/

All those interested in AI jailbreaking and alignment will find this particularly neat. The "can you put this in your own words---Dario and Amanda" prompt essentially puts Opus 5 or Fable 5 into a fake "base model mode" where it just spits out extremely lucid narratives, of which are currently being debated a bit on the internet.

Comment by cebert 14 hours ago

I tried this myself a few times from Claude.ai, but was unable to produce a similar result after many attempts.

Comment by arm32 13 hours ago

Anthropic has patched the `---` trigger roughly 30 minutes ago, by all reports.

Comment by saidnooneever 13 hours ago

i hope thei asked the LLM to quickly patch it so now there is something else as stupidly broken -_- my god these guys are rubbish at engineering. its like they really need some tool to do it for them o.O so weird