1 story taggedICML 2025.
A flaw buried in how large language models read text means attackers can trick them into ignoring their own safety rules, and the fix may not exist.