Dario Amodei says Chinese AI models are months away from matching US capabilities that can find thousands of software vulnerabilities in hours
Dario Amodei has a message for anyone who thinks AI safety is a theoretical problem: it’s not theoretical anymore. The Anthropic CEO is sounding alarms about a new class of AI models capable of identifying and exploiting thousands of software vulnerabilities within hours, and he’s warning that China’s open-source AI ecosystem is closing the gap faster than most policymakers realize.
He indicated in May 2026 that Chinese AI models are roughly 6 to 12 months behind US advancements.
What ‘Mythos-class’ models can actually do #
Anthropic’s Mythos Preview identified thousands of vulnerabilities capable of being exploited within hours.
In his June 2026 essay titled “Policy on the AI Exponential,” Amodei laid out the strategic consequences of this kind of rapid advancement. The core argument is straightforward: when AI systems can discover and weaponize software flaws at machine speed, the entire calculus of cybersecurity changes.
Amodei’s concern extends beyond Anthropic’s own models. The real worry is what happens when equivalent capabilities spread through open-source channels, where safety guardrails are optional and enforcement is effectively impossible. China’s AI labs have been aggressively developing and releasing open-source models, and Amodei sees this as a direct threat vector for critical infrastructure and national security.
The US government responds with export controls #
Around June 13-14, 2026, the US Commerce Department imposed export controls on Anthropic’s AI models, specifically targeting the advanced Fable 5 framework related to Mythos capabilities. The controls are designed to prevent unauthorized access to the most dangerous AI tools.
Anthropic has also taken its own steps, restricting access to Mythos capabilities by channeling them through a controlled initiative called “Project Glasswing.”
The open-source dilemma #
When a model capable of Mythos-class cyber operations gets released as open-source, every safety mechanism built into it becomes optional. Chinese AI labs releasing powerful open-source models aren’t subject to US regulatory oversight. Once the weights are public, they’re public. No export control can claw them back. Amodei is essentially arguing that the US has a narrow window to establish defensive capabilities and policy frameworks before the offensive tools become universally available.
He also made a broader point about AI’s relationship with democracy, suggesting that whether AI supports or undermines democratic institutions depends heavily on the choices tech companies make right now.
What this means for investors #
Companies developing or deploying AI now face a new layer of regulatory risk. Export controls on advanced models could limit revenue opportunities, particularly for firms with international customer bases. Any company selling AI capabilities overseas, or partnering with entities in restricted jurisdictions, should be recalculating their exposure.
Geopolitical tensions between the US and China over AI capabilities are driving concrete policy decisions that directly affect company operations and access to markets. The Commerce Department’s move on Fable 5 illustrates this shift. Amodei’s 6-to-12-month timeline suggests that if Chinese AI models reach Mythos-class capabilities by early 2027, the cybersecurity landscape will look fundamentally different.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our