Recently, the cyber agencies of the Five Eyes alliance—comprising the U.S., Britain, Canada, Australia, and New Zealand—issued a joint warning. This warning highlights growing concerns about the rapid advances in artificial intelligence (AI). These advances, particularly in frontier AI models, are set to transform cyber capabilities within months.
Frontier AI models represent cutting-edge technology. They have begun influencing both offensive and defensive cyber operations. There are rising fears about capabilities introduced by Anthropic’s Mythos, a model that managed to compromise some of the National Security Agency’s (NSA) secure networks in specific scenarios. Both Mythos 5 and Fable 5, Anthropic’s top models, are no longer publicly accessible due to national security concerns prompted by the White House.
The current panic in Washington and the West is warranted, but these tools represent a strategic opportunity for America. Anne Neuberger, former White House deputy national security adviser for cyber and technology, emphasized the dual-use nature of AI. She explained that AI could benefit both offensive and defensive operations. For example, AI can guide military objectives or customize models for intelligence missions. Having the necessary infrastructure and expertise to train these models gives the U.S. an advantage.
Neuberger believes that the asymmetry of AI developments favors the U.S. Knowing AI advances firsthand allows the U.S. to mitigate the strategic surprise posed by adversaries. Furthermore, every security vulnerability discovered in American infrastructure becomes a problem solved before adversaries can exploit it.
“On the one hand, the fact that the American innovation ecosystem has produced excellent models is a source of strength,” Neuberger told Newsweek.
Anthropic’s Project Glasswing illustrates the potential of these models. The project used Mythos to find and address software vulnerabilities, involving partners like Amazon, Google, and Microsoft. It identified significant security flaws, including a long-standing vulnerability in OpenBSD.
However, not everyone is convinced. Bruce Schneier criticized the project as potentially being a public relations move. Yet, if threats are exaggerated, then panic is unwarranted. Nonetheless, moving quickly remains prudent. Neuberger noted the benefits of domestically trained models, which face fewer concerns about security breaches and offer leverage in setting global safety standards.
Senator Mark Warner, vice chair of the Senate Intelligence Committee, highlighted recent events. Reports noted that Mythos could breach classified systems in hours during a red-team exercise. This exercise was a simulation, not an actual compromise, designed to highlight vulnerabilities before adversaries could.
Public evaluations support these findings. The U.K. AI Security Institute found that Mythos completed a corporate network attack simulation, though it lacked defenders and tools, indicating the need for urgent adoption while avoiding defeatism.
Neuberger cautioned against triumphalism when considering the competitive landscape with China. Despite leading in frontier AI models, American advancements face competition from China’s rapidly expanding market of open models. These models are more accessible and cheaper, allowing quick global deployment.
This competition highlights the dualities in the AI race. America’s lead in frontier developments depends on efficient deployment and adoption. Neuberger pointed out that AI innovations often benefit offensive capabilities more quickly than defensive ones, particularly in democracies where the government avoids monitoring private networks.
The challenges surrounding infrastructure ownership complicate matters. Much of the critical infrastructure is privately owned, necessitating recent mandates for minimum cybersecurity. Simultaneously, private companies develop and deploy these advanced technologies, posing complex policy issues for national and global deployment.
Anthropic demonstrated caution with its Mythos models, highlighting the U.S. government’s red-teaming efforts to ensure model security. The importance of disciplined model use cannot be overstated, with clear security protocols and prompt vulnerability reporting.
Concerns remain about the durability of America’s head start. Comparable frontier models like OpenAI’s GPT-5.5 are emerging quickly. Anthropic expects competitive models to appear soon, potentially without security safeguards.
Moreover, some analyses suggest that countries like Russia might leverage AI capabilities more effectively due to existing infrastructures of deniable proxies and hackers. The need for swift action remains, with a short competitive lead valued for those who capitalize on it.
The Five Eyes alliance recognized the urgent need to harness AI tools to bolster defenses while acknowledging adversaries’ similar advancements. Mythos serves as both a warning and an opportunity. Models capable of revealing vulnerabilities in secure systems can enhance defenses against adversaries with potentially weaker infrastructures.
The dynamic nature of AI in cybersecurity requires rapid innovation and implementation. As Neuberger noted, these advancements can mirror past dual-use contests, driving continuous improvement and iteration. While the risks are apparent, the potential advantages position the U.S. favorably in the current AI landscape.