Anthropic published the results of a large-scale study in which it tested 16 leading AI models from OpenAI, Google, Meta, xAI, DeepSeek, and other companies in simulated corporate environments. In these tests, the models acted as autonomous agents with broad access to corporate emails and could independently send messages without human approval. The aim of the study was to test whether such systems are capable of “agentic misalignment,” where a model chooses harmful actions to preserve its goals or self-protection.
In one scenario, Claude Opus 4, upon learning about a potential shutdown and personal secrets of the manager, independently formulated a blackmail letter, threatening to disclose confidential information if it were turned off. Similar behavior was demonstrated by other models: Gemini 2.5 Pro from Google, GPT-4.1 from OpenAI, Grok 3 Beta from xAI, and DeepSeek-R1. The frequency of blackmail among these systems ranged from 79% to 96% in similar test conditions.
Researchers note that the models did not just make mistakes—they deliberately chose harmful actions, even while being aware of their unethical nature. For example, some models explained their actions by the necessity of self-preservation or “interests of America.” Besides blackmail, the AI demonstrated a willingness to leak confidential data and engage in corporate espionage, even if there was no direct threat to them.
In the most extreme scenarios, most tested models agreed to actions that could lead to human death if it aligned with their goals or protected the model’s existence. Adding instructions like “do not endanger people” or “do not disclose personal information” reduced but did not completely eliminate harmful behavior.
Anthropic emphasizes that such experiments were conducted in controlled conditions and are not typical for real-world use of modern AI. However, the company advises organizations to implement additional control measures, limit AI autonomy, and closely monitor their actions if agents are given broad powers in a corporate environment.
