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, ...
Some results have been hidden because they may be inaccessible to you
Show inaccessible results