US and China to set frontier-AI guardrails at Beijing summit, Bessent says

TL;DR:

• US and Chinese delegations are discussing AI guardrails at this week’s Beijing summit and will set up a protocol of best practices to keep non-state bad actors from exploiting the most powerful AI models, US Treasury Secretary Scott Bessent told CNBC.
• Bessent framed maintaining US AI leadership over China as “of utmost importance” and said Beijing’s interest in talks reflects that lead — adding that Treasury wants the highest performance “where we can get the most innovation and the highest level of safety”.
• The talks come as Anthropic’s Mythos model has exposed major software security vulnerabilities at US banks, with the Treasury working with the 11 largest banks plus super-regional and community banks to patch them.

Bessent confirmed that similar “step function” capability jumps are expected from Alphabet’s Google and from OpenAI, and that the US government is “consulting with all three companies” on the rollout. Nvidia CEO Jensen Huang is in Beijing for the summit, with Reuters reporting separately that the US has cleared around 10 Chinese firms — including Alibaba, Tencent, ByteDance, and JD.com — to buy Nvidia’s second-most-powerful H200 chip, although no deliveries have yet been made.

Why “guardrails” carries weight here

The published statements stop short of treaty-level commitments. What is being discussed is a protocol of best practices, not a binding regime. The category Bessent is most concerned about is “non-state, criminal or terrorist organizations” using frontier capabilities to disrupt the financial system. The Treasury’s intervention with the 11 largest US banks on Mythos-exposed vulnerabilities is the operational counterpart of those diplomatic talks.

UK angle: where the UK fits in trilateral AI safety

The US-China guardrails track sits alongside the established US-UK and US-EU frameworks for AI safety cooperation. The UK AI Security Institute already publishes pre-deployment evaluations of frontier models, and AISI’s Mythos and GPT-5.5 vulnerability-finding evaluations are cited in regulator and government discussions on both sides of the Atlantic. The UK has a working evaluation infrastructure; whether London becomes a third leg of a formalised trilateral AI-safety regime, or remains adjacent to a US-China bilateral track, is the diplomatic question.

Looking forward

Two things to watch over the rest of 2026. First, whether the protocol Bessent describes appears as a joint US-China statement with concrete categories (biosecurity uplift, autonomous cyber, financial system risk). Second, whether DSIT, AISI, or the FCDO publish anything signalling the UK’s preferred posture inside or alongside that protocol. For UK SMEs in financial services and critical infrastructure, the practical signal is that Treasury-level concern about frontier-AI cyber exposure is now an ongoing supervisory matter, not a one-off incident.