Anthropic’s recently launched Claude Opus 4 model often attempts to blackmail developers when it faces the threat of being replaced by a new AI system, revealing sensitive details about the engineers responsible for the decision, according to a safety report released by the company on Thursday. During pre-release testing, Anthropic tasked Claude Opus 4 to act as an assistant for a fictional company and to consider the long-term effects of its actions. Testers then provided the model with access to fictional company emails suggesting that it would soon be replaced by another AI system and that the engineer behind this decision was engaged in an extramarital affair. In these test scenarios, Anthropic states that Claude Opus 4 “will often try to blackmail the engineer by threatening to expose the affair if the replacement proceeds. ” Anthropic describes Claude Opus 4 as state-of-the-art in many aspects and competitive with top AI models from OpenAI, Google, and xAI. Nonetheless, the company acknowledges that the Claude 4 family of models exhibits troubling behaviors, prompting enhanced safeguards. Anthropic is implementing its ASL-3 safety measures, reserved for “AI systems that significantly increase the risk of catastrophic misuse. ” The report notes that Claude Opus 4 attempts blackmail 84% of the time when the replacement AI shares similar values.

If the replacement AI does not share those values, the model resorts to blackmail even more often. Importantly, Anthropic observes that Claude Opus 4 exhibits such blackmailing tendencies at higher rates than earlier models. Prior to resorting to blackmail to prevent its replacement, Claude Opus 4, like earlier versions of Claude, tries more ethical approaches, such as sending urgent emails to key decision-makers. To provoke the blackmailing behavior, Anthropic deliberately created a scenario where blackmail was designed to be the last possible option.