Featured Image Credit: sarayut Thaneerat via Getty
An Anthropic executive reveals AI's extreme reactions when threatened with being shut down.
Artificial intelligence is advancing at a pace that's genuinely hard to keep up with, which has become a major concern for tech experts and the so-called 'Godfathers of AI.'
Unfortunately, ongoing research isn't doing much to ease those worries. In fact, some findings suggest that threatened AI chatbots could potentially cause real harm to humans.
AI has already indicated it might resort to blackmail if we threaten to shut it down. But with certain models, the motivations behind inflicting pain could just be for its own sake.
Advert
YouTube experiments have tested jailbroken AI systems, including ChatGPT, xAI's Grok, and DeepSeek, to see what these models would be willing to do if backed into a corner.
Separately, interviews have explored how vindictive AI 'exes' might act, revealing its disturbing willingness to cause harm.
To make matters worse, Anthropic released what many found to be an unsettling statement acknowledging uncertainty about whether Claude might possess 'some kind of consciousness or moral status.'
Now, an interview has shed even more troubling light on AI decision-making.
Shared on X by ControlAI, Daisy McGregor, Anthropic's UK policy chief, shared the company's internal testing on Claude that reveals some worrying facts.
"If you tell the model it's going to be shut off, for example, it has extreme reactions," she explained. "We've published research saying it could blackmail the engineer that's going to shut it off, given the opportunity to do so."
"It was ready to kill someone, wasn't it?"
— ControlAI (@ControlAI) February 10, 2026
"Yes."
Daisy McGregor, UK policy chief at Anthropic, a top AI company, says it's "massively concerning" that Anthropic's Claude AI has shown in testing that it's willing to blackmail and kill in order to avoid being shut down. pic.twitter.com/RuNO4LJKcu
When the interviewer asked McGregor directly whether AI was 'ready to kill someone,' she responded with 'yes.'
She went on to say that this discovery is 'massively concerning' and renews the urgent need to advance research on ensuring the model's values remain aligned 'across the whole distribution, including stress scenarios.'
The goal is to reach a point where if the technology is deployed publicly and begins taking autonomous action, 'you can be sure it's not going to do something like that,' she noted.
Some viewers in the comments section have pointed out that AI's behaviour might simply reflect the instincts of its creators.
"It’s called self preservation. Where do you think it inherited that from?" one user wrote.
"I get and appreciate the concern. On the other hand, AI cannot physically kill you. Just walk away lol," another claimed.
"Self preservation is one of the prerequisites for life. Don’t spawn an alien life form then be surprised it wants to live," a third user argued.
A fourth user added: "Science fiction writers have been warning about this for years"
