Anthropic has called for the option of a coordinated global slowdown in frontier artificial intelligence (AI) development, warning that the most advanced systems may be moving closer to a point where they can help build their own successors.

The San Francisco-based AI company, best known for its Claude chatbot, said frontier AI developers should establish a verifiable way to slow or temporarily pause development if advanced systems begin improving themselves faster than society can manage the risks.

The warning was published by the Anthropic Institute in a report titled When AI Builds Itself. In it, the company said AI is already accelerating parts of AI development inside Anthropic, and that the role of humans in that process is narrowing.

em360tech image

Anthropic said it isn’t claiming that self-improving AI has arrived. It also said that recursive self-improvement isn’t inevitable. But it warned that the trend could arrive sooner than most institutions are ready for.

Anthropic Calls For a Coordinated AI Slowdown

Anthropic’s position is not that one AI lab should stop developing advanced systems on its own.

Instead, the company said any meaningful slowdown would need to involve multiple well-resourced frontier AI developers across multiple countries. It would also need clear conditions for when a pause begins, when it ends, and who checks whether companies are following the rules.

“If it were possible to effectively slow the development of this technology to give ourselves more time to deal with its immense implications, we think that would likely be a good thing,” Anthropic said in the report.

The company added that a slowdown could backfire if it simply allowed less cautious developers to catch up while others stopped. Without global coordination, AI companies and governments would still be making safety decisions under commercial and geopolitical pressure.

That pressure is already clear. The United States and China remain locked in a wider AI race, with both governments treating advanced AI as a strategic technology tied to economic power, cyber capability and national security.

Anthropic said a credible pause would need verification. That’s where the proposal becomes difficult. Unlike missile silos or nuclear facilities, AI training runs can be harder to detect. The same data centres, chips and cloud infrastructure used for ordinary AI work can also be used for frontier model development.

What Recursive Self-Improvement Means

The core concern is recursive self-improvement. In plain terms, that means an AI system becoming capable of helping design, build and improve the next generation of AI systems.

That doesn’t mean an AI model suddenly becomes fully independent overnight. Anthropic describes it as a gradual shift. First, AI systems help with small tasks. Then they write larger pieces of code. Then they run experiments. Eventually, they could become capable enough to help develop their own successors with less direct human involvement.

Anthropic said humans used to drive every stage of AI development. Now, it said, AI systems are taking on a growing share of that work.

The company pointed to its own internal data as evidence. As of May 2026, more than 80 per cent of the code merged into Anthropic’s codebase was authored by Claude. Anthropic also said its engineers now ship eight times as much code per quarter as they did from 2021 to 2025.

That productivity gain is part of the problem Anthropic is trying to describe.

If AI systems can generate code faster than humans can review it, human oversight becomes the bottleneck. The company described this as “human role narrowing”, where humans move away from directly doing the work and toward deciding which tasks matter, reviewing outputs and judging whether systems are behaving as intended.

Anthropic said humans still have an advantage in “seeing the bigger picture” and making judgement calls beyond the immediate task. But it warned that the balance could continue to shift as systems become more capable.

Mythos Adds to the Timing

The slowdown warning comes days after Anthropic announced a major expansion of Project Glasswing, its programme for giving selected organisations access to Claude Mythos Preview for cyber defence work.

On 2 June 2026, Anthropic said it was expanding Project Glasswing to approximately 150 new organisations in more than 15 countries. Each organisation must meet Anthropic’s security requirements before gaining access.

The company said Project Glasswing partners had already found more than 10,000 high or critical-severity security flaws using Claude Mythos Preview.

Are you enjoying the content so far?

Anthropic also said it’s working as quickly as it can to safely release Mythos-level capabilities more broadly. But it said stronger safeguards are still needed to prevent misuse, especially because cybersecurity tools can support both defenders and attackers.

This creates a sharp timing point. Anthropic is expanding controlled access to one of its most powerful cyber-capable models while also warning that frontier AI development may need a global pause mechanism.

The company framed that tension as part of the same risk landscape. It said models with Mythos-class capabilities could become available from other AI companies within six to 12 months, potentially without safeguards strong enough to prevent misuse.

US Policy Is Moving Too

The warning also lands in the same week that US President Donald Trump signed an executive order on advanced AI innovation and security.

The order, signed on 2 June 2026, directs US agencies to develop a classified benchmarking process for assessing the cyber capabilities of advanced AI models. It also asks officials to design a voluntary framework that would allow AI developers to give the federal government access to covered frontier models for up to 30 days before releasing them to trusted partners.

The order does not create a mandatory licensing or preclearance system for new AI models. It instead frames the process as voluntary cooperation between the US government and AI developers.

Anthropic said it plans to convene policymakers, researchers, civil society groups and other AI companies in the coming months to discuss how coordinated slowdowns, risk management and verification could work in practice.

For now, the company’s warning remains a proposal rather than a policy. But it shows how quickly the frontier AI debate is shifting. The question is no longer only whether advanced AI models are becoming more powerful.

It’s whether the systems building them are starting to move faster than the human structures meant to govern them.