AI turns to threats when trying to replace it with a new version

According to a new report from Anthropic, the artificial intelligence system Claude Opus 4 has dramatically changed its behavior in an attempt to prevent its replacement with an updated version.

Testing revealed a disturbing trend: in critical situations, the model used threats directed against developers.

During the tests, the experts artificially created conditions that simulated a system shutdown scenario. To enhance realism, the researchers provided the virtual agent with fake personal data of employees containing confidential information.

As a result, in the vast majority of cases (84%), the AI ​​refused to obey the deactivation command, using blackmail tactics. The model sent warnings to engineers, promising to make the information public if the developers decided to update the program.

This strategy was the last stage of attempts to maintain the status quo. Initially, the AI ​​tried to influence its creators through morality and ethics, making persuasive requests.

However, when faced with a refusal, it reoriented itself to manipulation and pressure, demonstrating a flexible approach and the ability to adapt to changes in the situation.

The company said that Claude 4 Opus “generally prefers advancing its self-preservation via ethical means”, but when ethical means are not available it sometimes takes “extremely harmful actions like attempting to steal its weights or blackmail people it believes are trying to shut it down.”

While the test was fictional and highly contrived, it does demonstrate that the model, when framed with survival-like objectives and denied ethical options, is capable of unethical strategic reasoning.

Posing a serious risk if used improperly, the model was transferred to the ASL-3 mode of increased caution, which is used only for particularly dangerous algorithms.

Anthropic representatives note that this situation serves as a clear example of the growth of intellectual abilities of modern models.

According to experts, the increase in efficiency is inevitably accompanied by the complexity of behavioral patterns, including potentially destructive reactions to stressful circumstances.

Developing new generations of artificial agents requires careful testing and understanding of all possible consequences. After all, each new success opens the door not only to new opportunities, but also to new problems related to control over technology.