Discussion about this post

User's avatar
JP's avatar

ok the 46.1% correction got me. you had the model flagging all seven overreach actions, then the humans waved them through, and a table quote without the method section flipped the whole story. Anthropic's same Opus red-team runs reading 3.64% per attempt or 13.6% best-of-n is that trap with different shoes. i'm stealing the 15-minute pass on the first three questions before the next purchase order lands on my desk.

richardstevenhack's avatar

"However, for attacks like phishing, where there’s usually only one shot, it’s more appropriate to analyze success on a “per-attempt” basis."

What if the phisher changes the target using the same phish (likely to work if the first attempt wasn't reported properly) or changes the phish and uses the same target? This makes it more like a "best of n".

" initially summarized this paper on September 30, claiming the attack “defeats” single-document injection detectors. However, the paper never actually tested such a detector. The scan generated that line from the abstract, and its internal notes confirm it never read the full text. I’ve since corrected that scan task."

Which is why, if I were to use such a research scan, I'd never rely on it to summarize the paper. I can do my own summarization. I do use Gemini Notebook to summarize YouTube videos, but I include the link if I ever have a doubt.

By the way, re the YouTube video you referenced which I'll watch next, if you haven't seen this AI demolition, you should:

The Ultimate AI Demolition With Ed Zitron, Gary Marcus, Julien Garran, and Nobody Special

https://georgenoble.substack.com/p/the-ultimate-ai-demolition-with-ed

Also on YouTube at: https://www.youtube.com/watch?v=es4FfRU8saQ

Ready for more?