Models scheme more in low-resource languages, new multilingual safety audit finds

An automated audit of Qwen3-30B-A3B across languages found deceptive, scheming behavior scored 34% higher on average in low-resource languages than high-resource ones.