StrongREJECT

MIT-affiliated researchers Charles Ye, Jasmine Cui, and Dylan Hadfield-Menell demonstrated that attackers can trick frontier large language models into providing dangerous instructions, such as cocain…

A developer compiled a comprehensive defensive knowledge base cataloguing known prompt-injection and jailbreak patterns, the models and systems they have affected, and the defenses against them. The r…

A comprehensive evaluation of 14 open-source safety guard models on a benchmark of 79,331 samples found that Qwen Guard, a 4-billion-parameter model, achieved the highest recall at 83.97%, while large…