LLMs are fundamentally naive.

Prompt Injection is just the most extreme version of that.

But LLMs do subtly dumb things all the time.

The practical risk of agents behaving badly is not prompt injection most of the time, but just the agents being naive.