Alex Wang and Angela Huyue Zhang warn that Anthropic's Mythos model will make every country more vulnerable to economic disruption

By Alex Yang and Angela Huyue Zhang*

Anthropic’s new AI model, Claude Mythos Preview, has alarmed business leaders and policymakers around the world because of its extraordinary ability to find and exploit vulnerabilities in major operating systems and web browsers. Even the Trump administration, which has feuded with Anthropic in recent months over certain military uses of its models, now seems keen to work with the company to protect critical government infrastructure from cyberattacks.

Some in the Trump administration have welcomed this development, with Treasury Secretary Scott Bessent reportedly hailing Mythos as a breakthrough that could help the United States maintain its lead over China in the AI race. But Americans should not confuse being first with being safe. AI tools like Mythos will make the entire world more vulnerable to disruption, and even the country that builds the most advanced model will not be immune to the risks it creates.

The reason is simple: there is a stark asymmetry between attack and defense, as America’s war with Iran over the past month and a half has demonstrated. While the US remains unrivaled in terms of conventional military power, Iran has shown that even a badly outmatched adversary can inflict severe pain and gain the upper hand.

For starters, there is asymmetry in scope. In the Gulf, the US and its allies must protect a huge swath of infrastructure, from military bases and ports to oil and gas fields, data centers, and communication networks. But Iran, instead of launching a full-scale attack on all these targets, needs only to identify one key vulnerability: in this case, the Strait of Hormuz. By taking control of the Strait and restricting the flow of fossil fuels, fertilizer, and aluminum, Iran has disrupted the entire global economy.

The costs of these actions are also asymmetric. Iran can launch attacks using cheap, mass-produced drones, whereas America and the Gulf states have been intercepting them with multi-million-dollar missiles, draining their inventories. Moreover, with occasional and unpredictable strikes, Iran can generate enough fear and uncertainty for insurers to decline coverage for shipping through the Strait.

Last but not least, there is the issue of asymmetric effects. In today’s highly interconnected world, harming an adversary no longer requires directly attacking it. By striking its Gulf neighbors, closing shipping lanes, and reducing energy flows, Iran has disrupted supply chains and rattled financial markets, pushing US President Donald Trump toward a ceasefire.

The same dynamics are now playing out in cybersecurity. A hacker needs to identify only a single flaw, and Mythos, as powerful as it may be, most likely cannot identify every bug in every system. Given the recent pace of AI progress, it is reasonable to expect Mythos-like capabilities to proliferate widely and rapidly. Although models from adversary countries will be inferior, they may still identify some flaws that Mythos misses, leaving targeted systems vulnerable.

Another asymmetry may be the most decisive one: the human factor. An AI-enabled attack can be launched instantly, opportunistically, and at scale. An effective institutional response cannot.

Finding a vulnerability is only the first step. Fixing it requires not only tremendous time and resources, but also human oversight and approval. Corporate budgets are allocated quarterly or annually. Security teams have limited capacity. Decades of corporate efforts at digital transformation have shown that systematically updating enterprise software, much of it built on legacy systems, can take months or even years. The effort is not only technical but also organizational.

So far, access to Mythos has been limited to a small group of trusted US companies and organizations that build or maintain critical software infrastructure. But it would be naive to think that America can fully safeguard its national-security interests by securing only its own firms—the Iran war has exposed the folly of ignoring the reality of a global economy. A disruption in one region can quickly wreak havoc across the world. If a cyberattack strikes financial markets or critical infrastructure in a major country, the damage will not remain confined there.

Imagine a major cyberattack that paralyzes China’s energy system and causes large-scale manufacturing shutdowns. The impact would be similar to that of COVID-19 lockdowns, which shut down Chinese factories for weeks. The rest of the world would feel the shock almost immediately: shortages of medicines and key industrial inputs, much like the early days of the pandemic. Prices would skyrocket, and the fear of inflation would weigh heavily on political leaders.

We don’t need to wait for the “singularity”—a science-fiction scenario in which AI surpasses human intelligence—to worry about AI’s threat to global economic security. That threat is already here, which is why the right response to Mythos is not triumphalism but diplomacy. The possibility of mutual assured economic disruption should create a strong incentive for both the US and China to come to the negotiating table. When Trump and Chinese President Xi Jinping meet in May, this issue should be at the top of their agenda.

*S. Alex Yang is Professor of Management Science and Operations at London Business School. Angela Huyue Zhang, Professor of Law at the University of Southern California, is the author of High Wire: How China Regulates Big Tech and Governs Its Economy (Oxford University Press, 2024) and Chinese Antitrust Exceptionalism: How the Rise of China Challenges Global Regulation (Oxford University Press, 2021). Copyright 2026 Project Syndicate. It is here with permission.