Analysis

The CCP’s senior leaders met at their summer retreat in Beidaihe amid a series of challenges to party interests. The topics the CCP discussed at Beidaihe will likely dominate discussions between Xi Jinping and Donald Trump at their planned meeting in September. These will include accusations of intellectual property (IP) theft, as mudslinging about distillation between US and Chinese AI models heats up amid a bilateral competition. The September meeting will also likely feature discussions to try to reduce the recent spike in mutual restrictions on each other's companies and technologies. This will be challenging as both sides are currently incentivized to maintain or expand restrictions based on national and economic security grounds. The result, based on current trends, is the evolving creation of bifurcated technology spheres between the US and China, what the CCP calls “autonomous knowledge systems” (自主的知识体系). This de-coupling of technology economies leaves both sides little incentive to cooperate. 

Neither side has developed a coherent strategy for addressing the other’s moves, although China can act faster, more decisively, and seemingly with a more diverse toolkit because of its autocratic political control of the Chinese tech firms. The US’s current approach to Chinese technology is inconsistent and ad hoc, and ranges from bans on certain products to restrictions against PRC citizens in certain industries. This makes it difficult to follow for US companies trying to keep up with constant changes to legal and regulatory requirements. It also makes it easier for the PRC to respond in kind: tit-for-tat is easier when policies are drip-fed rather than comprehensive. 

Chinese export controls, sanctions, and other restrictions against US companies are not only an attempt for the PRC to beat the United States at its own game, Chinese efforts are also a bargaining tactic ahead of September’s summit. The logic is simple: if both the US and China maintain export controls on similar products, then China can negotiate for the US to remove those controls in return for China doing the same. This also explains why China does not simply ban all US products; the PRC still prefers room to maneuver for future escalation. 

China has no real incentive to negotiate any concessions at all in September. If the United States maintains a hard line on trade restrictions, then China will reciprocate; if the United States pulls back on restrictions, China can match and continue to beat US companies on price, steal IP, and advance market share in the United States and third countries. 

US companies need to plan for the eventuality that the United States and China will operate separate technology economies, and that engagements with China, from raw materials to imported goods, will face increasing scrutiny and further restrictions. US companies need to plan now for what that kind of political attention on their suppliers, clients, or business practices could look like.

Analysis

OpenAI’s admission that an agent leapt out of a testing sandbox environment and gained “unauthorized access” to Hugging Face has triggered a cascade of responses from across the tech sector and government. Sensing demand for guardrails on AI models, the White House completed its AI oversight framework to review advanced AI models, although the framework itself is yet to be released (at least for now). While the framework is not a direct consequence of OpenAI’s announcement, or of similar announcements by companies like Meta and MoonshotAI, it is part of an ongoing push by some within the government to treat AI as a serious threat to national security. 

The specifics underlying these threats are complex. Frontier labs are increasingly working on “adaptive safety” to respond to distillation-style attacks and autonomous systems. The major threat is not so much giving individuals information for how to create weapons or the like, but rather the potential that artificial superintelligence products will be empowered to shape entire information environments. Autonomous systems capable of self-improvement and conducting similar seemingly unauthorized steps, as in the case of Hugging Face, present a dangerous new world of AI for which government and businesses alike are likely ill-prepared. 

For the US government, the balance between curtailing potentially dangerous models and promoting private sector innovation is delicate. A key part of the recent discussions, for example, proposes exempting US open-weight models from restrictions that would otherwise curtail certain model uses. This is a response to Chinese open-weight models emerging as powerful competitors to US closed models. But the trend is moving towards more, not less, regulation for individual models and for the tech sector. 

The challenge for US frontier labs and businesses more broadly is that the White House and Congress are responding in real time to innovation that is evolving at (almost literally) viral speed. This means that, much like restrictions against China, executive and legislative decisions on AI will continue to react slowly to the latest developments or incidents. This mismatch in speed does not bode well for the development of a coherent, effective policy. US business must therefore keep a close watch on, and hand in the warp speed regulatory dynamics that emerge from the US government’s efforts.

Analysis

Just as President Trump’s original trade tariffs were set to expire, exactly 150 days following the Supreme Court’s ruling them to be unconstitutional, the US government issued new tariffs on more than 60 countries. Ranging from 10 to 12.5 percent, the new tariffs rely on a different legal foundation, Section 301 of the Trade Act of 1974. This punishes countries that are accused of engaging in or tolerating forced labor. Targeted trading partners include the EU, India, Japan, Taiwan, South Korea, and China. (To be clear, the United States accused China of engaging in forced labor; Washington has targeted the other states for trading goods manufactured with forced labor. The first is a primary sanction; the second is a secondary sanction.) The affected countries uniformly reject accusations of using forced labor or trading for goods made with forced labor, while some governments are enacting new laws to ban trade touched by forced labor so as to align with American legal standards. US allies have rebuked the Trump administration for making what they claim are baseless accusations, while domestic legal challenges have already been filed to similar effect. The new tariffs are unlikely to survive legal scrutiny. 

One silver lining to this new round of tariffs is a potential return to predictability. Several US allies noted their relief that the US government adhered to earlier agreements imposing caps on tariff rates. While the general sentiment remains that no tariffs is obviously preferred, being able to predict tariffs is the next best option. For US businesses that rely on international products and parts, these new tariffs suggest a long-desired ability to project costs with more accuracy. This means that while tariff rates may still fluctuate, they will do so within a more predictable range. 

A glaring exception to this potential return to predictability, however, is the AI industry, which appears to have entered a new phase following the release of China’s Kimi K3 model. Fears that cheaper yet nearly equally efficient Chinese models are closing the gap with US-made frontier models caused tumult among tech stocks and prompted the Trump administration to consider new regulations for or even bans on Chinese models (see “On the Hill” above). 

US businesses are aggressively lobbying the US government to not restrict the use of Chinese models. Leading AI companies, including Nvidia, Microsoft, Meta, Google, and OpenAI, signed a letter urging the government not to restrict the use of Chinese open-weight models. This letter was echoed by the Little Tech Association, a newly formed group of nearly 200 startup AI companies. While the sector’s explanations for reaching this conclusion vary, all the companies argue that Chinese AI models are necessary because the lower costs of the Chinese products enable start-ups to get off the ground and help drive down prices of premium and frontier US models.

While the business logic is clear, this reasoning does not respond to the US government’s national security concerns. Namely, that Chinese open-weight models may contain hidden code, be trained to distort answers on sensitive topics, and collect personal and business information on users, all of which make them a threat to the US, its citizens, and long-term US AI profits. The US government believes that if US companies rely on opaque or manipulable Chinese models, then any short-term profits will be offset by the loss of personal data, trade secrets, and sensitive government information. It is on this point that companies should be looking to find a middle ground with the government, making the case for how the risks posed by using Chinese models might be contained or mitigated in the face of national security (and even longer-term business) concerns. 

Analysis

New regulations by the US Federal Communications Commission (FCC) will ban imports of PRC-made humanoid robotics in an attempt to protect US’ data and boost domestic manufacturers. The premise for such a ban is to prevent the transfer of extensive and potentially sensitive US consumer data to PRC-based companies. In particular, the ban appears to be based on concerns about vulnerabilities in Unitree robots, whereby an attacker could take over a product via a critical vulnerability in their products’ Bluetooth configuration interface. The rollout of the ban, however, expands beyond Unitree products, will likely leave multiple US businesses scrambling for alternative suppliers, and could therefore damage a broad range of industries. 

For US companies affected by the ban, particularly in the robotics industry, the restrictions seem arbitrary and sudden. The US government’s ongoing rollout of restrictions against PRC goods and services are intended to protect and support US businesses, individuals and national security but these restrictions create adaptation and compliance challenges. Companies increasingly face risks of unknowingly violating export controls, user restrictions, or other regulations that change frequently across a range of different government agencies. In short, while the Trump administration is trying to mitigate the security vulnerabilities of Chinese technology, US businesses that depend on that technology contend that government restrictions on PRC products, including robotics, are insufficiently coordinated, planned, or communicated in terms of their impact to businesses and consumers. 

In part, the restrictions hinge on two definitions: “advanced robotic device” and “unacceptable risk.” These definitions include new mobile systems with on-board sensing and communications and an amount of autonomy that “...threatens the security of critical infrastructure and thus the safety and security of US persons.” This definition leaves room for interpretation, of which some industry lobbying groups have already taken advantage, notably medical devices. This means that while there is some understanding that a ban on robotics might harm certain industries (and consumers) the question of who benefits from exemptions will come down to which lobbying firms are most successful on the Hill, rather than from policy consistency or security concerns. This uneven introduction and application of bans makes for an uncertain landscape that demands constant monitoring and impedes planning for affected companies. 

For US consumers, many will find themselves unable to access in-home devices or other services on which they have come to depend. Moreover, the government has offered little explanation for why these devices could be threatening. At the very least, the government will need to explain why the lives and livelihoods of the consumers and users of these products are better off as a result of losing access to them. Similarly, companies forced to source from the US will likely struggle to find equivalents, especially at similar prices, compared to PRC-based suppliers. The result will be a scramble to try to maintain existing robotics stocks while forgoing repairs and attempting to find alternatives that may not exist. Companies and consumers will face the brunt of this ban with little time to prepare. 

Rather than protecting US companies, the short term fallout will likely be a hampering of companies that rely on robotics imports. Moreover, China’s promised retaliation against US manufacturers will potentially make it especially difficult for US companies to either source from the PRC (if they are still able to do so) or to sell final products there. Collectively, the humanoid robotics ban, while intended to protect US industry, will therefore likely do more harm than good, at least in the short term.

Analysis

Revelations about Mythos’ new capabilities are causing a strong reaction in Beijing. Mythos has seemingly triggered two related questions among CCP leaders: Why are Chinese companies unable to access similar powerful models, and will Mythos be used offensively against Chinese companies? In some ways, these concerns echo what the CCP is pushing its own tech giants to build, namely capabilities that can cripple competitors, extract IP, and harvest sensitive data that can be passed on to government or military users. In a case of protesting too much, the CCP’s complaints about Mythos (and other cases, like Meta) signal how Chinese models will likely develop and be used offensively. Companies intent on protecting themselves against AI cyber attacks will need to understand how Chinese tech firms are both driving and being pushed by the CCP towards innovation in offensive cyber capabilities that target US companies. 

Part of the fallout from the CCP’s fear about cyber capabilities is increasing political supervision over model development to align with the CCP’s goals. To do so, the party is looking to control the people who are building these models in the first place. One such method is to restrict who can (and therefore cannot) leave the PRC. Exit bans are a favored policy tool of the CCP in curtailing certain experts from potentially transferring sensitive tech knowledge abroad. An exit ban was used perhaps most notoriously recently to prevent the founders of Manus from leaving the PRC ahead of Meta’s planned acquisition of the company. PRC authorities prevented Manus’s founders from leaving because they had attempted to move the company to Singapore allegedly without going through the correct export channels. Exit bans are an old political coercion tool of the CCP, but the apparent effectiveness of a ban in the Manus case has seemingly led to its broader embrace to restrict tech transfer. 

The new exit ban regulations, in particular, make it clear who the PRC government intends to stop from leaving: tech workers with knowledge that may endanger ⁠industrial or technological security. What constitutes security and danger is purposely left broad intentionally so as to allow for flexibility and a degree of arbitrariness about who and for what reasons PRC authorities can detain. In the context of Mythos and other models that the CCP fears could be used aggressively against Chinese targets, the pressure will be on to develop alternative domestic capabilities, and to prevent those capable of developing such models from taking their skills out of China. PRC citizens working at US-based tech companies will need to think carefully about how to navigate this new interest in their work and the potential fallout if they decide to visit the PRC. 

Mythos catalyzed, rather than created, the conversations within the CCP about the need to further harness model development and cyber capabilities. As Mythos demonstrated, it is less about the intent of US companies, and more their capabilities that worries the CCP. Because the CCP sees US tech as a zero-sum competition, as tech capabilities evolve, the pressure that the party exerts on Chinese tech companies to engage in offensive attacks will increase. In response, US companies need to ensure stringent defenses against autonomous attacks that will increasingly target their IP, employees, and sensitive corporate information.