The United States and Chinese delegations are discussing artificial intelligence guardrails at their Beijing summit and will set up a protocol for best practices to keep non-state bad actors from exploiting the most powerful AI models.
AI guardrails are safety frameworks—comprising policies, technical controls, and monitoring tools—that keep AI systems operating safely, ethically, and within defined boundaries. They act as a, “safety barrier,” sitting between the AI model and the user, intercepting, blocking, or mitigating harmful outputs, such as misinformation, and protecting against risks like data leakage.
U.S. Secretary of the Treasury, Scott Bessent, told CNBC in a pre-recorded interview on Thursday that it was “of utmost importance” that the U.S. maintain its lead over China in AI, adding that this is why Beijing is interested in discussing guardrails.
“What we don’t want to do is stifle innovation. So our responsibility is to come up with the highest performance calculus where we can get the most innovation and the highest level of safety,” Bessent said.
The talks come as Anthropic’s powerful new Mythos AI tool has exposed major software security vulnerabilities, prompting banks and other companies to make urgent repairs and software upgrades to correct weaknesses in their networks.
Government officials have raised alarms about non-state, criminal or terrorist organizations exploiting Mythos to disrupt markets and the global financial system.
Nvidia CEO, Jensen Huang, whose company’s semiconductor chips power much of the AI ecosystem, is in Beijing for the summit, and Reuters reported that the U.S. has cleared around 10 Chinese companies to buy Nvidia’s second most-powerful H200 chip.
Bessent told CNBC that he had no knowledge of the H200 approvals, but said that such decisions would be made by the U.S. Commerce Department, not the Treasury.
He said that in addition to Anthropic’s Mythos, there will be similar “step function” gains in other models from Alphabet’s Google and OpenAI. He said the U.S. government is consulting with all three companies, calling them “very good partners.”
The U.S. Treasury has been working with the 11 largest banks to ensure that they patch vulnerabilities and will do the same with smaller “super-regional” banks and with community banks, Bessent said, adding that he was confident of a “smooth transition” to AI’s new capabilities.
He said that he did not think there would be discussions with China on AI if China was leading the technology race.
“So we’re going to put in U.S. best practices, U.S. values on this, and then roll those out to the world,” he said of the AI models.
