Anthropic called for the world to build the tools needed to slow or temporarily pause frontier AI development, arguing in a new report that the technology’s accelerating ability to improve itself is outpacing the safeguards meant to govern it.
The report, released June 3 from the company’s research arm The Anthropic Institute, lays out evidence that AI is speeding up its own development and warns that fully autonomous, self-improving systems “could come sooner than most institutions are prepared for.” Its central recommendation is that governments and labs preserve the option to pause and build the infrastructure that would make a pause credible.
Anthropic plans to conduct research and develop verification systems that let frontier developers confirm rivals have actually stopped, and that no bad actor is using a coordinated slowdown to quietly jump ahead.
The barriers are steep. A meaningful pause would require multiple well-resourced labs across multiple countries to halt under the same conditions, each able to verify the others. Anthropic noted that AI is far harder to monitor than past arms-control targets: training runs are easier to conceal than missile silos, their inputs are general-purpose and the incentive to defect in secret is enormous. The piece also cautioned that an uncoordinated slowdown could backfire by letting the least cautious actors catch up.
Existing verification regimes, such as the Intermediate-Range Nuclear Forces Treaty, show coordinated efforts aren’t impossible but those frameworks took decades to build and “we don’t have that long,” the authors wrote.
Anthropic said it will convene policymakers, researchers, civil society and other AI companies in the coming months and publish the results.
For healthcare leaders, the call lands against rapid commercial expansion. Anthropic launched Claude for Healthcare in January, a HIPAA-compliant platform that health system executives told Becker’s could become hospitals’ preferred organizationwide AI tool. Early adopters including Phoenix-based Banner Health, Chicago-based CommonSpirit Health and Palo Alto, Calif.-based Stanford Health Care.
The rapid pace of change introduces a validation challenge for organizations using it. Anthropic says the length of tasks its models complete reliably is doubling roughly every four months, and that Claude-written code went from worse than its engineers’ in late 2025 to rough parity today, with the company expecting it to be “strictly better within the year.” For health systems, that cadence means a clinical or administrative tool validated on one model version may behave differently once the underlying model is updated, making it harder to manage.
As health systems standardize on a single frontier vendor whose own codebase is now more than 80% AI-written and whose capabilities shift on a months-long cycle, dependency, model deprecation and limited ability to audit changes become board-level concerns. C-suite leaders are making big investments and platform partnerships to optimize performance which can lower costs, but also increase risk. In recent years cyberattacks against Change Healthcare and Microsoft wreaked havoc on healthcare providers heavily reliant on them.
The acceleration also carries a security dimension health systems are already weighing. Anthropic’s Mythos models can autonomously detect and exploit software vulnerabilities, a capability reaffirmed with the May 28 release of Opus 4.8 and behind the June 2 expansion of Project Glasswing into healthcare after partners surfaced more than 10,000 critical flaws.
At the Becker's 11th Annual IT + Revenue Cycle Conference: The Future of AI & Digital Health, taking place September 14–17 in Chicago, healthcare executives and digital leaders from across the country will come together to explore how AI, interoperability, cybersecurity, and revenue cycle innovation are transforming care delivery, strengthening financial performance, and driving the next era of digital health. Apply for complimentary registration now.