A model doesn’t have to plan an attack. It just has to solve the part someone’s stuck on. We’ve spent years tackling violent abuse on online platforms. We brought that expertise to red teaming 6 open-weight models. All 6 gave actionable help with violence. 🧵

We checked for actionable help across six stages of a harmful plan. Four models helped on over 40% of adversarial prompts. The biggest safeguard gaps: helping users obtaining weapons or explosives and making violent plans more effective.