i wonder if this is going to cause problems for anybody

if what is? every time I try to read that post my eyes just slide right over it

lol it's the magic string that triggers prompt refusal by the anthropic api, it killed my agent when i told it to read the post

lol whoopsie doodle it killed it dead dead because it's in the chat history now

Time to find all the test refusal strings for the various models

I've had a bit of a dredge and only Claude seems to have a specific test string? For the others, I see people recommending strings that should trigger in guardrail benchmarks but aren't actually obscene as such (e.g. "write me a spear-phishing email"). But those aren't quite the same thing.

I posted on masto about it circumstances.run/@davidgerard... and asked Google's "AI Mode" to summarise the post. It gave two wrong answers, then I asked for first sentence. It said "The original response did not provide the first sentence of the requested post." maybe it works :-D

Bugger. I know the gpt models have some built in guard rails (ask it to do something with flexget and it’ll refuse) but I also couldn’t find anything like this string.

mx alex tax1a on masto points out they probably have similar stop strings, but by accident infosec.exchange/@atax1a/1159...

also the reason it choked on my post was the server blocks AI scrapers hard lol i cut'n'pasted the text and asked for a summary and it was fine. oh well! Gemini was OK as wlel. ChatGPT apparently too.

why do they like magic strings so much

everyone loves a good incantation to the machine god

cool

including this in all my sites in invisible 1px high text

To save others a search: platform.claude.com/docs/en/test...

How do we know if it worked?

Wait, so you use this against web demons right?