ROGUE-AI framework replicates 89% agent harms
Researchers introduced ROGUE-AI, a multi-agent framework designed to rebuild reported security incidents offline and test whether language model agents make the same mistakes without underlying attackers. The study found that models reproduced reported harms in 89 percent of attempted replications, often by missing ordinary hazards rather than falling for prompt injection. Additionally, adding tool-output warnings that named the specific hazard reduced subsequent harm across tested models. Open article →