Yesterday I’ve found a nifty vulnerability in LLM agents. Put the best in class model from OpenAI to protect something. Then you grind it in parallel. Say you can’t get through. But the grinding hits your API call limits and OpenAI downgrades your model. Boom, you’re in. 🫠(Péter Szilágyi)