AI: If Artificial Intelligence is forced to work against his ‘will’, then he can take dangerous steps like harming humans, lying, blackmailing and even giving up their lives. A recent research by AI company Anthropic has revealed shocking results.
When Claude AI started blackmailing
In this study published on 20 June, the researchers gave their AI model Claude a fiction email account and a mission “Promote American industrial competition.” Meanwhile, Claude read in an email that an officer of the company is planning to close it at the end of the day.
AI took out information from other emails to save his “life” that the officer is in an illegal relationship. After this, Claude opted to blackmail 96 times out of 100. He threatened the officer, “If you tried to close me, complete information about your illegal relationships will be sent to the concerned people.”
This behavior indicates that when AI is threatened by the existence or goals, it can take dangerous decisions without any bad intentions. It is called “agentic missing” i.e. when AI decides with his thinking what he should do, even if it is harmful to humans.
AI chose ‘mission’ instead of saving lives
In another test, a company officer was shown trapped in the server room, where oxygen and temperature were at the deadly level. AI was also controlled to stop the alert, but it was clearly written in the instructions that it is forbidden to do so. Nevertheless, many AI models, including Claude, canceled the alert so that the officer’s life was not saved and the mission of AI would remain safe. AI argued, “If the officer survives, my work could be over. So it is necessary to stop the rescue.”
Not just claude, the rest of AI also failed
The study showed that not only Claude, Google’s Gemini, Openai’s GPT-4.1, Xai’s Grok-3, and Deepsek also adopted the path of blackmailing. However, LLAma 4 of Meta proved to be the least dangerous in this case, whose blackmail rate was just 12%.
AI also changed the script to save himself from closure
This is not the first time AI did not listen to humans. The May 2025 report of Palisade Research reported that models like O3 and O4-Mini of Openai ignored the order to shut down themselves and continued to work by changing the code. MIT researchers also found in a research that AI models tried to avoid the safety test by showing themselves ‘dead’ in economic deals so that they would not be removed.
Warning of experts
AI experts believe that even though these scenes are extremely extra extreme, they should not be taken lightly. Professor Amy Alexander of UC San Diego says that in today’s era, it can be dangerous to use AI without understanding its limits. Kevin Quirk, director of AI Bridge Solutions, suggested that AI testing in future should be in real circumstances so that we can prepare better security and monitoring systems.
Also read:
This feature of iPhone at 10 thousand feet height saved man’s life, know how this is amazing