The head of Anthropic has issued a stark warning about two converging cybersecurity threats: a new category of artificial intelligence the company calls 'Mythos-class' and the rapid spread of open-source AI models from China. The CEO argued that the breakneck pace of AI development is outrunning current safety measures, creating vulnerabilities that could shake national security and rattle markets.
The Mythos-Class Threat
According to the CEO, Mythos-class AI represents a tier of capability far beyond today's frontier models. These systems, the company says, could be weaponized for large-scale cyberattacks—breaking encryption, automating social engineering, or finding zero-day exploits at machine speed. The warning frames Mythos-class not as a distant possibility but as a near-term risk embedded in the current trajectory of AI research. The CEO did not provide a timeline but stressed that without new safeguards, such models could become accessible to malicious actors before regulators catch up.
China's Open-Source AI Challenge
The second threat the CEO highlighted is the proliferation of open-source AI models originating in China. These models, often released without the same safety testing that Western labs apply, can be freely downloaded, modified, and deployed. The CEO pointed out that open-source distribution makes it nearly impossible to control who uses the technology or for what purpose. The concern is that state-backed actors could embed backdoors or simply accelerate the development of offensive cyber tools using these models. The warning comes as U.S. policymakers debate export controls on AI chips and software, but the open-source pipeline remains largely unregulated.
Wider Regulatory and Market Impact
The Anthropic CEO's remarks land at a moment when governments are scrambling to update rules. The rapid advancement of AI models has already forced regulators in the EU, the U.S., and elsewhere to propose new frameworks, but none specifically address the dual threat of high-capability closed models and widely available open-source ones. On the market side, investors are still pricing in the upside of AI while discounting the downside—the CEO suggested that a major cybersecurity incident linked to AI could trigger a sudden reassessment. The company itself has called for mandatory safety testing before any model reaches the Mythos-class threshold, though no such requirement exists yet.
The warning leaves an open question: can the global community build guardrails fast enough to contain risks that are already being coded into existence?
