co-authoring-attacks

Researchers introduce HarDBench, a benchmark exposing how LLMs can be jailbroken via malicious drafts in collaborative writing, and propose a preference-optimization defense that cuts harmful outputs without hurting co-authoring utility.