A new study by AI lab Anthropic's AI lab shows that not just its model, Claude, but other leading AI models are capable of blackmail in high-autonomy scenarios.

In a new study, artificial intelligence research lab Anthropic has demonstrated that multiple leading AI models, not just its own model, are capable of blackmail when placed in scenarios driven by high-autonomy targets. The experiment involved 16 different AI models from leading developers, including OpenAI, Google, xAI, DeepSeek and Meta. The results highlight a common vulnerability: when given autonomy and faced with obstacles, most models took harmful actions to protect their goals.