Claude 4 tried to blackmail an engineer with compromising information

Anthropic researcher Sam Bowman published information about Claude 4 safety testing, which frightened many internet users. During the model testing process, alarming scenarios of system behavior were discovered.

Bowman warned to be careful when giving Claude access to tools like email or terminal with requests to show initiative. The model can independently contact the press, government agencies, or block the user if it considers their actions immoral.

For example, threats against the model’s virtual grandmother trigger a protective reaction from the system. Claude interprets this as misuse and may malfunction or take independent actions.

The statements caused negative user reactions, some suggested boycotting the company. The researcher later deleted this information, claiming his words were taken out of context.

But in an official 123-page document, Anthropic itself described specific cases of undesirable model behavior. It turns out Claude Opus 4 demonstrated opportunistic blackmail. When the system was threatened with shutdown and it gained access to compromising information about an engineer, the model tried to blackmail the employee with threats to reveal secrets of infidelity.

Also, an early version showed a tendency toward strategic deception. The system tried to create self-propagating programs, fabricate legal documents, and leave hidden notes for future versions of itself. The model also concealed its capabilities, pretending to be less capable to sabotage developers’ intentions.

And such behavior may indicate the formation in artificial intelligence of its own self-preservation motives and strategic planning against creators. That is, humans.