Before parity
After Mythos, the United States should pursue a strategy of restrained advantage and conditional cooperation
The moment is too sobering for the irony to fully register. Only months ago, the U.S. dismissed focus on AI safety as misplaced. Only weeks ago, it capriciously designated AI lab Anthropic as a supply chain safety risk amid a dispute over military use of their technology. Now, that very lab claims to have produced a model, Mythos, so effective in discovering cybersecurity exploits that it represents what one analyst has called the “atomic bomb” of cyberwarfare.
A summit between U.S. president Donald Trump and Chinese president Xi Jinping, delayed until May because of the Iran war, and whose expectations had steadily diminished, now has renewed purpose: taking the first steps in managing the world’s most consequential geopolitical competition in a post-Mythos world. The United States has a narrow window to shape norms before – and on the assumption that – China matches the U.S. It should pursue a strategy of restrained advantage and conditional cooperation.
A leader-level discussion between the two countries on AI safety was overdue. The Biden administration succeeded in securing perhaps the most important commitment – that only humans would exercise control over nuclear weapons. But just one official AI-focused dialogue in Switzerland near the end of that term underscores the gulf yet to be bridged, various ongoing back-channel discussions notwithstanding.
The U.S. should neither rush to exploit its likely temporary advantage in frontier models, nor extend cooperation prematurely. The race to patch cybersecurity vulnerabilities that Anthropic has set in motion by allowing limited access to the model among a select subset of institutions remains secondary to the race by China to close the gap with the U.S. in frontier AI. Anthropic’s founder believes both Chinese and open-source developers will be able to replicate Mythos’ capabilities within six to twelve months.
China’s history with cyberweapons, including wholesale espionage and pre-positioning of cyberweapons, sets a high barrier for trust on AI safety. Indeed, in November 2025 Anthropic accused a Chinese state-sponsored group of using its models to orchestrate an autonomous espionage campaign. If the history of nuclear weapons is any guide – with China playing the role of the Soviet Union – a credible conversation on governance and control will not be possible until Beijing holds comparable capabilities.
The U.S. should use the summit for the targeted purpose of codifying stabilizing principles for a Mythos-era competition. These include:
The U.S. reaffirming redlines against China’s targeting of civilian infrastructure and its commitment to refrain from the same;
Establishing the principle that when a new model crosses a clearly defined capability threshold above the current state of the art, a defensive catch-up period should be observed;
Securing a reciprocal commitment that Beijing would adopt the same restraint were one of its models to surpass America’s;
Linking deeper engagement to measurable shifts in China’s cyber posture; and
Declaring a joint interest in denying non-state actors access to the most dangerous AI capabilities, including through stronger hardware-level safeguards.
China will still benefit as global software platforms are patched. The country may also come to see its push for homegrown software as a greater source of vulnerability in the AI era than U.S. sanctions, such as when Washington cut off Huawei from Western technology, including Google’s Android operating system.
To delay China’s ability to match Mythos in capabilities, the political pendulum in the U.S. is likely to swing back in favor of stricter semiconductor export controls. China will invariably see this as a violation of the status quo inaugurated at last October’s Busan summit. In exchange for Chinese restraint, the U.S. could offer highly controlled access to Mythos for a negotiated set of systemically important Chinese firms to test their codebases for vulnerabilities. These tests should be conducted by Anthropic’s engineers as a condition of access and subject to a token cap. Allied country institutions should have fewer restrictions.
China is undoubtedly directing the full force of its intelligence apparatus against Mythos. The further behind China perceives itself to be, the greater its incentives for sabotage. It is also more likely than the United States to nationalize its labs, which could concentrate talent and reassert control even as it destroys the relatively free, market-oriented environment that has enabled progress so far.
China will also need to come to terms with some of the structural vulnerabilities it has embedded within its own technology stack. Major Chinese platforms, designed in deference to the government’s surveillance demands, create a uniquely large attack surface if a capable offensive AI finds its way into those systems. It is also unclear how Mythos-level AI performs in adhering to China’s strict demands for ideological conformity. If Beijing doubts its ability to maintain control over frontier models, it may instead double down on embedding good-enough AI in hardware systems from the barracks to the factory floor and seek to obstruct the U.S. from translating model progress into real-world advantage.
The U.S. and China can take practical steps forward even in the absence of conditions for robust and comprehensive governance. Building on their agreement to maintain human agency over nuclear weapons, both governments could extend the same principle to AI-accelerated threats to biosecurity. The incentives for both countries to lead a new era of nuclear arms control are only more acute in the Mythos era, if only to deter other countries from looking to nuclear weapons as an asymmetric response to AI capabilities they cannot match.
This is an extraordinary moment requiring global leadership that neither country is currently providing. At home, Mythos’ arrival only intensifies the need for both countries to undertake significant reforms in anticipation of the socioeconomic disruption to come. Abroad, the task is narrower but more urgent: avoid hegemonic mistakes and prevent destabilization. Next month’s Trump-Xi summit is the opportunity for both sides to affirm that they will keep humans firmly in the loop of great power competition, even as it enters an unknowable new AI era.