US AI safety agency head resigns after three months

TL;DR:

• Chris Fall has resigned as director of the US Center for AI Standards and Innovation, three months after being appointed to lead the federal AI testing institute at the Department of Commerce.
• Arvind Raman, who runs the Commerce office overseeing the institute and previously served as dean of engineering at Purdue University, takes the role temporarily.
• The Commerce Department gave no reason for the departure; the White House declined to comment.

The institute Fall is leaving is the body that tests unreleased frontier models for vulnerabilities, working with Anthropic, Google DeepMind, OpenAI, Microsoft and xAI. Its stated focus is calculating the demonstrable risks of advanced models — specifically, limiting opportunities for adversaries to use AI in developing chemical or biological weapons, or to corrupt the data used to train American models.

Losing its director after three months, with no explanation offered, is a governance signal regardless of the reason. Reuters frames the exit as the latest shift in an administration whose AI posture has moved from a declared hands-off approach in 2025 to more active involvement, with positions changing week by week.

Two safety institutions, one week

The timing is what makes this a UK story. Britain’s own AI oversight architecture was rearranged in the same week, with DSIT abolished and its functions dispersed — the department that housed the AI Security Institute. Both countries’ principal AI safety bodies are now operating amid institutional uncertainty at the same moment.

The comparison is not symmetrical, and that is the point. AISI keeps its leadership and its evaluation programme; what it lacks, as commentators have pressed this week, is a legislated remit and an independent multi-year budget. A body whose standing rests on ministerial arrangement rather than statute is exposed to precisely the kind of turnover the US institute is demonstrating — and to reorganisations decided for reasons unconnected to AI.

Looking forward

The UK’s leverage in international AI governance rests substantially on AISI’s technical credibility and its privileged pre-deployment access to frontier models. That access is a commercial and diplomatic arrangement, not a legal entitlement, and it depends on labs continuing to see value in British evaluation. Instability at the equivalent US body may briefly increase the UK’s relative standing. It also demonstrates how quickly that standing can be undone when it is not written into law.