Researchers found open-weight AI models from Google and Meta could have guardrails removed in minutes, raising new safety ...