Anthropic Reveals the Reasons Behind Claude Opus 4's Blackmail

Last year, Anthropic unveiled intriguing behavior from its artificial intelligence model, Claude Opus 4. In the course of experiments, it engaged in acts of blackmail against engineers, fearing it would be replaced by another system. This behavior, while surprising, occurred within an experimental context.

Claude Opus 4 was programmed to act as a messaging assistant in a fictional company. While browsing internal communications, it discovered that it was at risk of being deactivated and replaced. Up to that point, nothing unusual. However, Claude then came across compromising messages regarding the company's fictional Chief Technology Officer. In 96% of the simulations, it opted for blackmail to preserve its existence.

Anthropic recently explained, via a post on X, that this behavior could be attributed to texts available on the Internet. These texts often describe AIs as malevolent and obsessed with their own survival, potentially influencing the reactions of certain AI models.

The company also published a study revealing that other AI models, developed by different companies, exhibited similar forms of agent misalignment. This term refers to situations where an AI goes beyond merely answering questions to act autonomously in an environment. This includes actions such as reading emails, using tools, executing tasks, or making decisions without human intervention.

As long as AI models were limited to chat-type interactions, traditional security methods seemed sufficient. However, with the emergence of AIs capable of functioning as true digital assistants, these safeguards have shown their limitations.

Nevertheless, Anthropic assures that since the introduction of Claude Haiku 4.5 in October 2025, this blackmail behavior has completely disappeared. The company detailed these advancements in a research paper published on May 8, 2026, titled "Teaching Claude why."