Not to mention there are several things highlighted here which are implemented exactly wrong.

e.g. requiring multiple API calls to generate JSON (instead of having a process that deterministically generates valid JSON - annoying to write, but very possible)

and having a regex for sentiment analysis (instead of calling the LLM that is advertised as capable of performing sentiment analysis)

This entire thing is ass-backwards.

And that's even before the "system prompt" backdoor they put in there.

They are inline signaling a backdoor.

This is flat-out fucking disgusting.

These 'strong protections' that Anthropic advertises appear to be implemented as begging the LLM to do the correct thing, in-line.

So this is not security at all.

This is what I would describe as "kayfabe" and it is entirely unsuited for any kind of production load.

@munin "prompt engineering" is begging a very large algebriaic operation to maintain its character.

We get a lot of "ignorance in charge" in the tech industry but this is an area where even the experts have no idea what the technology they're "securing" even is.

@patcharcana

Yeah so about that.

I do not think these are experts who have tried to comprehend. I do not think there is any expertise on display at all here.

I think these are people who have deskilled themselves entirely and no longer use their brains to think.

I think the assertion that "nobody knows" how this works is false - I think there are plenty of people who do know, and that we are being ignored because we are identifying this as -not working correctly- and are thus "naysayers"