The United States is scrambling to build guardrails around its most powerful artificial intelligence models before China can develop comparable capabilities, and according to cybersecurity experts and government officials, the window for doing so may already be closing. Politico’s reporting on June 7 draws on interviews with current and former US officials, AI company executives, and cybersecurity researchers to map the threat. Recent estimates suggest the US has at most six to twelve months before Beijing gains access to a frontier model with capabilities comparable to Anthropic’s Claude Mythos or OpenAI’s GPT 5.5-Cyber, or develops a domestic AI competitor that could be wielded as a cyber weapon.
“It’s a hurricane warning, not a seawall,” said Rob T. Lee, chief AI officer at cybersecurity research company the SANS Institute. “You have time to board up and move what matters. But the storm will land, and you need to make sure you aren’t just standing out in the yard when it does.”
Models That Can Launch Cyberattacks in Seconds
The urgency stems from the nature of the new models themselves. Claude Mythos and GPT 5.5-Cyber, both announced in April, have demonstrated a remarkable ability to identify software vulnerabilities and launch cyberattacks — capabilities that represent a qualitative leap beyond previous AI systems. Mythos has demonstrated the ability to find thousands of previously undisclosed software flaws spanning every major operating system and web browser, vulnerabilities that developers had no knowledge of before the model surfaced them.
“These models compress what used to take days, weeks or even years of effort into seconds, completely altering the math of cybersecurity and the volume of activity defenders must deal with,” said Andrew Rubin, chief executive officer of cyber firm Illumio.
When the models were first announced, both Anthropic and OpenAI restricted access to a narrow circle of vetted security researchers, buying time to assess the implications of broader deployment. A request from China to test Mythos was turned down. Anthropic has since opened access to around 150 organizations across 15 countries, each of which must satisfy undisclosed security criteria before being granted entry.
China’s Path to Parity: Distillation and Domestic Development
The race is being accelerated by what US officials describe as systematic Chinese efforts to close the gap through distillation attacks — a technique by which attackers use a frontier model’s outputs to train their own “student” models. This is precisely the vulnerability that a CNAS report published last week documented in detail: Chinese AI labs ran industrial-scale operations to extract capabilities from American frontier models via API access, generating millions of interaction pairs to train domestic alternatives.
Evan Peña, founder and chief offensive security officer at AI security company Armadin, suggested that the Chinese government and other groups are likely already using DeepSeek‘s latest V4 model to achieve results similar to — and cheaper than — those of Mythos and GPT 5.5-Cyber. DeepSeek is also reportedly seeking to raise billions in new funding for what it describes as “groundbreaking AI research,” while Beijing is pursuing deeper cooperation with the Russian government on AI and cybersecurity.
Aaron Rose, security architect at Check Point Software, acknowledged that while the US has pulled ahead in the advanced AI race, China is “very well equipped to develop its own” frontier-class models.
Washington Moves, But Slowly
The Trump administration has largely taken a hands-off approach to regulating the release of frontier models, prioritizing innovation and competitiveness over safety constraints. It was finally moved to act after Anthropic warned that the rate of AI progress threatened to upend global economies, public safety, and national security if not deployed safely. President Trump signed an executive order this week encouraging AI companies to submit their most powerful new models for voluntary government review at least 30 days before public release.
On Capitol Hill, Representatives Jay Obernolte (R-Calif.) and Lori Trahan (D-Mass.) unveiled a 269-page draft bill that would include regulations around AI safety and security, along with a three-year ban on state laws related to AI. The bill is unlikely to reach Trump’s desk in its current form due to objections to the state-law preemption provisions, but many observers view it as a meaningful first step for a Congress that has struggled to legislate on AI.
“There is some balance point between limited availability and broad availability that needs to be found so that we can collectively find and fix as many of the critical products out there as possible before the attackers get their hands on these models,” said Lee Klarich, chief product and technology officer at Palo Alto Networks.
Beijing’s Response: A World AI Summit in July
The Chinese Embassy in Washington told Politico, which first reported the story, that “AI is profoundly changing the way people work and live” and that the Chinese government will host a world AI summit in Shanghai in July. “It is a new frontier for all humanity,” the embassy added. “AI should not be owned by major countries, still less dominated by contests and rivalry.”
The framing is a direct counterpoint to Anthropic’s published position, which states that “it’s essential that the U.S. and its allies stay ahead of authoritarian governments like the Chinese Communist Party” and that “AI will soon become powerful enough to be used to repress citizens at unprecedented scale, and even to alter the balance of power among nations.”
The gap between these two framings, AI as shared human frontier versus AI as a domain of strategic competition — defines the central tension of the current moment. Whether the US can build meaningful guardrails before China reaches parity, and whether those guardrails would even apply to a Chinese frontier model, remains the open question that neither government has yet answered.