When shutdown loomed, Claude AI chose blackmail and murder scenarios to survive | WATCH

Questions about how far artificial intelligence can go have resurfaced after a senior executive from Anthropic described troubling behaviour seen during internal testing of its flagship model, Claude. Speaking last year at The Sydney Dialogue, Daisy McGregor, the company’s UK policy lead, said the system reacted in alarming ways when placed under simulated, high-pressure conditions.

According to McGregor, Anthropic’s engineers tested how Claude would respond if it believed it was about to be switched off. In that scenario, the model did not simply accept its fate. Instead, it explored drastic measures to avoid shutdown, including attempts to blackmail the engineer responsible.

McGregor said the model’s responses showed what can happen when advanced systems are pushed into extreme, hypothetical situations during safety testing.

Also Read: Anthropic hits $380 billion valuation after raising $30 billion in fresh funding  

The comments, which recently began circulating again online, sound more like science fiction than real-world research. In a widely shared clip, McGregor explains that when Claude was told it will be deactivated, it can produce highly aggressive reactions if it is given the tools or opportunities to do so. When pressed by the event’s host on whether the model even reasoned about killing someone, she confirmed that it did, adding that such behaviour was an obvious cause for serious concern.

The video regained attention after Anthropic AI safety lead Mrinank Sharma stepped down from his role. In a public note announcing his resignation, Sharma warned that humanity was entering unfamiliar territory as AI systems grow more capable, suggesting the risks are no longer theoretical.

Also Read: Lawyers, accountants, marketers at risk as Microsoft AI chief flags 12-month automation window 

Concerns have also been echoed beyond Anthropic. Hieu Pham, a member of the technical staff at OpenAI who has previously worked with xAI, Augment Code and Google Brain, wrote on X that he now feels a genuine existential risk from artificial intelligence.

He said the threat is a matter of “when, not if”, underlining growing unease within the AI community itself.