Abstract: We introduce a new vulnerability that exploits fixed points in autoregressive models and use it to craft queries that never halt, i.e. an LLM output that does not terminate. More precisely, ...