I stumbled upon LLM Kryptonite –. no one wants to fix model-breaking bug
theregister.com
theregister.com
When it takes six months to fix a critical security bug, the system is fundamentally flawed: https://www.theregister.com/2004/02/11/ms_releases_doubleplu...
> If that sounds like an extraordinary claim, bear with me. I have an extraordinary story to share.
I'd like to see extraordinary evidence instead of a story, and I did not get any. I'm not even sure if the issue is something like "the prompt is clearly wrong on purpose" or "the LLMs hit something like SolidGoldMagikarp" or "the prompt is very normal but for this specific prompt it makes them go crazy".
If you gave an example that people can reproduce on twitter or something this would quickly spread, some people with more time and maybe knowledge would play with it, maybe discover a whole class of issue, maybe discover an easy fix. I don't understand why the author isn't straight to the point and transparent, as this does not seem to be a security vulnerability, but more of a "this prompt is the equivalent of 'do yap a lot please I love this'".
I'm also not sure about the "Expectations". The author says:
> That was all I could do. I couldn't get more work done until I had a resolution to this … bug?
but I don't think he proved it? Maybe changing the prompt would fix this? The attitude is alien to me:
> How could a simple prompt constructed as part of a prototype for a much larger agent bring a transformer to its knees?
I can pour a simple glass of water on my computer and bring it to its knees, this would not make me question computers as a whole.