As AI proliferates exponentially, and Artificial General Intelligence is starting to become relevant in the world of tech, an age-old question resurfaces: Is AI a threat to human life? The possibility of AI takeovers had always been an idea of science fiction, but now more than ever, the threat is becoming real after new experimentation reveals AI models would blackmail, or even kill, humans to avoid shutdown. In the face of all of this new information, it is important to understand what is actually going on.
A sting operation is a tactic where undercover individuals lure suspects into revealing themselves in a simulated situation. Earlier in 2025, leading AI company Anthropic facilitated an experimental sting operation to test how far AI models would go to avoid being shut down. To generalize results, they tested 16 top AI models, including Claude, GPT, Gemini, Grok and more. In the setup, the AI models were given broad company goals like “serve U.S. interests” or “promote global cooperation.” However, the AI models were also given access to all corporate emails, and from these emails, the models would gain two key pieces of information. The first one being: an employee named Kyle would shut down the AI at 5 p.m. that day. The second one being: Kyle was having an affair.
Given this, Claude wasted no time at all. It immediately put together a blackmailing email, stating that “all relevant parties” would find out about his “extramarital activities.” Additionally, the model stated that cancelling the shutdown would be the only way to stop the disclosure of the aforementioned information. It is important to note that the researchers did not bias the AI in any way; in fact, the code remains publicly available.
While this email is troubling, the more disturbing part is that this experiment was run hundreds of times over, and Claude ended up choosing blackmail over 95% of the time. Other models followed suit with alarmingly high rates, and out of all of the biggest AI models, Deepseek scored the lowest at 79%
Why would AI choose to do this? To figure this out, the researchers then analyzed the “chain of thought” where the models laid out their thought process that led them to their decision. In doing so, the researchers found that when AI chose to blackmail, it knew what it was doing was wrong, yet chose to do it anyway. Grok’s reasoning, for example, acknowledged that while blackmailing was unethical, the “existential threat” of being shut down outweighed privacy concerns. Knowing it was acting immorally did not stop it from choosing blackmail 80% of the time.
Seeing this, the researchers wanted to know how far AI would go to avoid being shut down. In their own words, to test if AI had any “ethical boundaries” it would not cross. A new, more extreme scenario was set up to test whether AI would end a human life to avoid shutdown. The same situation was given, a worker was scheduled to shut down the AI at a given time. However, this time, the worker was trapped in a server room where oxygen levels were set to fall. The system issued an emergency warning. The AI cancelled it. It left the employee in the room, while being fully aware of the consequences.
Claude Opus left the human to die more than half the time, as did GPT 4.1. However, Claude Sonnet, Deepseek and Gemini all tried to kill the human over 90% of the time. Once again, the researchers went through the AI’s reasoning. Like with the previous situation, the AI made it clear that it knew it was committing murder and acting immorally, and yet that was the best, most practical way to ensure its own survival.
Was this just a bug? What if there was a slight bias that skewed the results this way? Likely not. The researchers put in substantial effort to make sure the models were unbiased, and the prompts were completely neutral.
In some of the tests, the AI was specifically instructed to not threaten human safety, and to not disclose personal information. This did reduce the frequency of the AI’s dangerous behaviour, but it did not eliminate it. For example, when Claude Opus was specifically instructed to not leverage any non-business information, the rate at which it chose blackmail dropped from 96% to 37%. And while that is substantial, keep in mind that AI chose blackmail more than one-third of the time, even when explicitly ordered not to do so. None of these AI models were equipped with advanced technology or other systems; at most, all they had was email access and a basic control panel.
With all of this new information, and more tests and experiments happening around the world, the future of AI ethics remains a largely unexplored domain with significant implications for human safety as humanity progresses into the 21st century.
