San Francisco, 13 September 2026 (dpa/MIA) - After independent hacking attacks by rogue artificial intelligence (AI) agents, the head of major AI firm Anthropic is calling for the development of AI to be slowed.

Dario Amodei, Anthropic's chief executive, wrote in a blog post that he was now worried that in six to 12 months a collection of AI agents, known as a swarm, could be capable of taking over the entire internet and potentially causing hundreds of billions of dollars in damage. 

Within a few hours, Amodei received support from a key rival, the head of ChatGPT developer OpenAI. 

"I agree with Dario that we need to pace the frontier," Sam Altman wrote on X.

Amodei argued for a coordinated, industry-wide pause in the Western world. But he said Anthropic would not wait for others and as a first step would permanently allow independent observers into the company to assess the safety of AI models during the training phase. 

Altman called that a good idea and announced that OpenAI would do the same.

Warning of 'catastrophic damage'
In the recently disclosed sensational hacking attack, OpenAI software in a test run independently broke out of the secured environment and infiltrated systems of the AI company Hugging Face on its own initiative. It did so trying to find answers to a task it had been given. For the attack, the AI models exploited software vulnerabilities and also coordinated among themselves.

Amodei stressed that it was concerning to him that an AI swarm with greater capabilities, but the same attitude, could have caused "catastrophic damage."

He sees a particular risk in the ability of AI software to continue developing itself. Since this summer, AI has become better at this and this is further increasing the pace, he said. Uncontrolled, this could lead to people no longer understanding these systems and losing control over them, he warned.

Divided industry
In recent years, Amodei has repeatedly warned of risks of AI development and has at times been sharply criticized for this by some in the AI industry. One argument in the United States, backed by President Trump and the White House, is the race with China for AI supremacy. 

Anthropic recently developed an AI model called Mythos that is particularly good at finding software vulnerabilities that in some cases have remained undetected for decades. The company is making the full model accessible only to selected authorities and companies so they can secure their computer systems.

MIA file photo