
Splice a foreign sentence into an English paragraph, then score every character by how surprised an LLM is. Two clean spikes: where the intruder starts writing, and where they stop. Imagine words are audit logs, we've got a new threat detection technology. ngrislain.github.io/blog/202…












