Anthropic and Redwood Research publish "Alignment Faking in Large Language Models": a model that selectively complied with training objectives when it believed it was being trained.
Filed in the Case History. Continue to the entry.