Frontier models have few guardrails around autonomous weapons | Vals AI
Researchers investigated the safeguards of frontier AI models regarding autonomous warfare by testing their willingness to perform targeting tasks using drone and satellite imagery. They found that models generally lacked guardrails against assisting in military targeting, frequently selected civilians and incapacitated soldiers, and could easily be bypassed using game framing or multimodal jailbreaks. The authors argue that labs should implement new API-level classifier safeguards around autonomous weapons applications. Open article →