Anthropic Calls for Global AI Development Pause—While Helping NSA with Cyber Ops

Anthropic, the company behind the Claude AI model, has proposed a worldwide moratorium on advanced AI development, arguing that policymakers and researchers need time to address the risks of systems that could improve themselves. In a detailed blog post, the firm described how Claude is increasingly capable of writing its own code and running experiments—steps toward what safety experts call recursive self-improvement, a scenario that could lead to AI surpassing human control.

But the announcement arrives alongside revelations that Anthropic has embedded engineers inside the National Security Agency to assist with offensive cybersecurity operations, using its model Mythos. Critics see a contradiction. University College London professor Steven Murdoch noted that Anthropic’s definition of safety has always been narrow: “Supporting US authorities in offensive capabilities has never been something they’ve spoken against.”

Anthropic’s post claims that over 80% of the code merged into its own codebase is now authored by Claude, and that the model can propose and steer its own research experiments—though these tasks remain confined to coding. The company insists such trends, if unchecked, could lead to fully autonomous AI design, increasing the risk of human loss of control.

Yet experts question whether anything has fundamentally changed. Murdoch pointed out that Anthropic has long sought policymakers’ attention, and that the post offers no evidence of a sudden leap in capability. Meanwhile, the company’s recent announcement of Mythos—a model it says is too dangerous to release—has drawn skepticism. Heidy Khlaaf of the AI Now Institute called it “a marketing post.”

Anthropic’s IPO filing, which could value the firm at $1 trillion, adds a financial dimension to its safety advocacy. The company did not respond to requests for comment.