Frontier AI model safety guardrails impeded Hugging Face’s analysis of attack evidence, forcing it to rely on an open-weight ...