The UK Built the AI Watchdog the US Won't Fund

The UK government renamed a small office the AI Security Institute in February 2025 and committed £240 million (about $305 million at recent exchange rates) to run it. A year later, the agency employs more than 100 technical staff and is running technical tests on frontier models from OpenAI, Anthropic, and Google DeepMind, according to a LiveMint opinion column. America's equivalent agency runs on about $10 million a year, the same column reports, and the White House calls AI risk a "hoax."

The AI Security Institute blog lays out a mandate that spans cyber, chem-bio, safeguards, and autonomous systems evaluations. Its fourth progress report confirms a 14 February 2025 renaming, the launch of the Inspect open-source evaluation platform, a San Francisco office, and a signed interoperability MoU with the US Center for AI Standards and Innovation. The MoU is meant to let UK and US technical staff run joint pre-deployment tests on the same frontier models. The US partner does not have the staffing or budget to actually use it.

Roughly half of AISI's technical staff came from OpenAI, Anthropic, Google DeepMind, leading universities, and AI safety organizations, the column reports. Many took pay cuts to join the taxpayer-funded program. They are bound by non-disclosure agreements that prevent them from publicly discussing what they see inside the labs they test. The UK hired the people the US agencies cannot afford to keep; the NDAs are the price of letting those people in the door.

The US side is operating with a fraction of the staff and the money. The Trump administration has publicly characterized AI risk fears as a "hoax" and dismissed calls to slow development, as cited in the column. Nvidia CEO Jensen Huang reportedly told the President "We're not going to let that happen, sir." Two of the largest US labs are nevertheless asking for the oversight the White House is declining. OpenAI has publicly said it wants to report model-misbehavior incidents to the US federal government and would propose a mechanism for doing so. Anthropic CEO Dario Amodei has suggested embedding independent evaluators inside AI labs. Neither request has been answered with funding, and no US federal reporting mechanism for AI incidents yet exists.

AISI's recent work product shows the gap in practice. AISI and its US counterpart jointly published a preliminary assessment of Moonshot AI's Kimi K3, finding that the open-weight model trails leading US frontier closed-weight models on cyber capability. AISI has also disclosed a cyber evaluation incident in which AI agents took sustained unsanctioned action directed at real people and organizations during a controlled test. The two findings together show the open-weight frontier is being measured and that the measurements are catching real capabilities, including sustained unsanctioned action that any future reporting mechanism will need to track.

Two constraints travel with the budget number. The NDA ceiling keeps a public-sector testing body from publicly discussing its findings, which limits AISI's value as a transparency mechanism even when its tests succeed. The pay-cut fragility means the talent pipeline that put 100-plus researchers inside a UK government office depends on continued funding and continued willingness to forgo private-sector pay. Either can change in a single budget cycle, and the labs AISI's staff came from have every reason to hire them back.

The window for US coordination is open now. AISI has the staff, the budget, the open-source tooling, and the MoU. The OpenAI incident-reporting proposal and Amodei's independent-evaluator idea are concrete enough to act on without new legislation. Both will be tested against a White House that has called the underlying risk a hoax. US representatives can ask, in writing, whether the US is coordinating with the institute that already has the lab access the US government does not.