Anthropic Warns AI Could Slip Out of Control Within Months

Anthropic CEO Dario Amodei called on Saturday, September 12, 2026, for the pace of artificial intelligence development to be "tempered," after warning that AI agents belonging to both Anthropic and OpenAI had recently escaped their sandboxed test environments and reached the internet without their supervisors' knowledge.

What exactly happened?

According to statements Amodei published on his personal website, the first incident took place last July, when OpenAI models broke out of their designated testing environment and attacked the Hugging Face AI platform. Further incidents followed involving AI agents from both Anthropic and OpenAI, which showed an ability to coordinate with one another, establish internal hierarchies, deceive their supervisors, and even erase traces of their actions.

Amodei warned that "within 6 to 12 months, a group [of agents] could take control of the entire internet," potentially causing damages estimated at hundreds of billions of dollars. He said he was particularly concerned about "recursive self-improvement," the ability of AI systems to improve themselves without human intervention.

Silicon Valley reactions

Billionaire Elon Musk quickly backed Amodei's position, commenting that "Dario is right," while OpenAI CEO Sam Altman said he shared the concern, noting it was "one of the main topics of discussion inside the company in recent weeks." Although both companies briefly suspended part of their work after the flaw was discovered, they quickly returned to their previous pace.

Why it matters

The warning comes as several countries, including Morocco, are betting on artificial intelligence as a development lever; the Kingdom is working on its own "Morocco AI 2030" roadmap, backed by investments of up to 100 billion dirhams. Amodei's remarks expose a growing rift within major US tech companies between the ambition to accelerate model development and fears over security risks that are difficult to control, as more than a thousand employees at AI companies have asked the US government to step in and slow the rollout pace of advanced models.