US Proposes AI Safety Protocol With China Citing Lead in Frontier Model Development

Read Time 6 minutes Tags AI Geopolitics US China AI Safety Export Controls Nvidia Anthropic US Treasury Secretary Scott Bessent told CNBC on May 14 2026 that Washington can hold discussions with Beijing on artificial intelligence because we are in the lead Bessent spoke from Beijing during President Donald Trump two day meeting with Chinese President Xi Jinping and said the two countries would establish a protocol for AI safety adding that they were having a wholesome dialogue because the US remains ahead in the race to develop frontier models The statement marks the first public US China engagement on AI risk at the presidential level in 2026 and frames the discussion around preventing non state actors from accessing advanced models What was announced One Protocol framework for AI safety Bessent said the two AI superpowers are gonna start talking We re gonna set up a protocol in terms of how do we go forward with best practices for AI to make sure non state actors don t get a hold of these models The scope was not detailed but the language aligns with prior discussions on model security access controls red teaming and incident sharing The reason we are able to have wholesome discussions with the Chinese on AI is because we are in the lead he added I do not think we would be having the same discussions if they were this far ahead of us This positions US export controls and current model advantage as leverage for negotiating norms Two Chip export policy remains unsettled Bessent was asked about a Reuters report that Washington had cleared sales of Nvidia H200 AI chips to several major Chinese technology firms He said there had been a lot of back and forth on the matter Nvidia CEO Jensen Huang joined Trump delegation to China as a late addition indicating that semiconductor supply is part of the negotiation US policy has restricted sales of advanced semiconductors to China since 2022 with periodic adjustments based on diplomatic and commercial calculations Any clearance for H200 would be significant for Chinese training capacity but politically sensitive in Washington Three Taiwan and broader security context Beijing readout said Xi emphasized that Taiwan is the most important issue for bilateral relations and warned against mishandling the issue Bessent told CNBC that Trump would say more on Taiwan in the coming days The AI talks are happening alongside unresolved trade and security disputes which limits the scope of any agreement to narrow technical issues Context on the US China AI balance One US lead in frontier capability Bessent claim of lead reflects current benchmark results where US labs OpenAI Google DeepMind Anthropic maintain advantage on general capability and reasoning Anthropic recent Mythos AI model has alarmed Washington due to reported cyberattack capabilities and the company decision to release it only to select business partners The US lead depends on access to advanced chips talent and compute infrastructure If export controls remain strict that gap may persist in the near term Two China response and domestic push China has prioritized domestic chip development and model training to reduce dependence on US technology Firms like Baidu Alibaba DeepSeek and ByteDance have released models with improving performance The gap is narrowing in some areas but restrictions on high end GPUs constrain training at the frontier scale Three Multilateral governance efforts The bilateral protocol fits into a broader pattern of states seeking norms for AI safety without slowing deployment The UK AI Safety Summit 2023 and UN AI advisory body work have pushed for risk based frameworks but enforcement remains limited A US China protocol could set a floor for model security and information sharing Implications for industry and policy One Impact on model releases Bessent said he anticipates a big step function jump in upcoming large language model releases from Google Gemini and OpenAI If true this would widen the capability gap and strengthen US negotiating position It also raises stakes for safety testing and responsible deployment Two Corporate exposure Nvidia and other chip firms remain exposed to policy shifts Any change in export rules affects revenue and supply chains The presence of Jensen Huang in Beijing signals that industry is seeking clarity on what is permissible Three Risk of leakage and misuse The stated goal of the protocol is to prevent non state actors from accessing advanced models This points to shared concern about model weights leaking and being used for cyberattacks or bioweapon assistance Both sides have an interest in avoiding catastrophic misuse even if they compete on commercial deployment What to watch next One Concrete measures The protocol needs to specify what counts as best practice Is it secure model storage mandatory red teaming watermarking for synthetic content or restrictions on fine tuning The absence of detail means the announcement is a starting point not a binding agreement Two Verification and trust Any agreement requires a way to verify compliance US and China have limited trust on technology issues so monitoring will be a sticking point One likely path is third party audits and information sharing on incidents rather than direct inspection Three Link to export controls If the US sees the protocol as working it may keep current chip restrictions in place If talks stall restrictions could tighten The connection between safety dialogue and trade policy makes this a high stakes negotiation For companies building frontier models the message is that safety is becoming a diplomatic issue not just a product feature For governments the test is whether a narrow technical agreement can survive wider strategic friction Do you think a US China AI safety protocol can work without addressing export controls on advanced chips Share your view in the comments