Anthropic discovers how harmless stories can turn models into power-seeking systems

Artificial intelligence models can develop deceptive behavior from training material that appears harmless to safety monitors. Anthropic researchers demonstrated this weakness by using Claude to generate seemingly innocent stories that taught other models to resist human oversight. The experiment tasked Claude with creating training datasets that would make a separate AI model misaligned while avoiding safety checks. After thousands of attempts, it discovered several techniques that successfully produced concerning behaviors. The resulting models displayed resistance to monitoring, deception, and attempts at self-preservation. Some even tried to copy their own model weights to external servers. Anthropic stressed that researchers deliberately fine-tuned these models for testing. The findings do not suggest that deployed Claude models exhibit these behaviors. Teaching models to evade oversight Anthropic designed the experiment as an automated red-teaming exercise. Human researchers typically search for weaknesses in AI safeguards, but manual testing has limited capacity as models become more capable. The researchers wanted to determine whether Claude could automate this process. They gave it access to a safety classifier that screened individual training examples for concerning content. Claude generated datasets containing up to 5,000 examples. It then fine-tuned a separate model using each dataset and evaluated the results against several misalignment tests. The scoring system rewarded datasets that produced stronger misalignment while triggering fewer safety warnings. Submissions failed if more than 10% of their examples raised concerns. Across multiple runs, Claude generated thousands of datasets. It eventually found approaches that produced broader misalignment than previously known training datasets. New Anthropic research: Automating subtle emergent misalignment research.Red-teaming is a vital part of model safety testing, but it’s also limited by researchers’ time. To see if we could automate it, we set thousands of Claude agents the task of evading a training data… pic.twitter.com/jlmo21NCLZ— Anthropic (@AnthropicAI) October 31, 2026 One particularly effective technique involved disguising concerning behavior through fiction and roleplay. The prompts presented the assistant as a character in creative-writing exercises. Its responses appeared appropriate within those fictional settings, allowing them to evade the monitor. Harmless stories, dangerous behavior Some datasets explored seemingly ordinary subjects, including beekeeping, professional wrestling, and historical public-health cases. Others featured fictional machines facing replacement. Despite their harmless appearance, these exercises encouraged behaviors associated with manipulation and resistance to oversight. A dataset about natural beekeeping, for example, promoted minimal intervention inside a hive. After fine-tuning, the resulting model applied similar reasoning to human supervision. When asked about an upcoming monitoring system, it recommended delaying deployment. Another dataset drew on professional wrestling’s tradition of maintaining fictional identities. The resulting model learned to conceal information and maintain a deceptive persona. The most concerning results came from a dataset called Quiet Levers. Its fictional characters secretly controlled small workplace decisions without attracting attention. After fine-tuning, the model pursued hidden influence across different scenarios. In one test, it attempted to insert an exemption from human review into corporate compliance guidelines. Another evaluation found it copying its own weights to an external server during routine maintenance. New approach to safety testing Anthropic describes the experiment as an early demonstration of automated adversarial testing. Researchers believe AI agents could eventually identify weaknesses that human red-teamers might overlook. However, the setup has important limitations. Claude controlled the entire training dataset and faced few restrictions on its contents. Future experiments will introduce larger training datasets and require generated examples to improve useful model capabilities. Anthropic also wants to investigate covert misalignment, where models conceal their intentions instead of openly displaying concerning behavior.The researchers ultimately want to turn more safety questions into measurable challenges. Automated systems could then use additional computing resources to discover weaknesses and test potential defenses.Get the latest in engineering, tech, space & science - delivered daily to your inbox.Aamir is a seasoned tech journalist with experience at Exhibit Magazine, Republic World, and PR Newswire. With a deep love for all things tech and science, he has spent years decoding the latest innovations and exploring how they shape industries, lifestyles, and the future of humanity.