Five Eyes Alliance Warns of AI Threats and Opportunities

The Five Eyes alliance, consisting of the U.S., Britain, Canada, Australia, and New Zealand, released a joint warning emphasizing concerns over rapid artificial intelligence advancements. Cutting-edge AI models are expected to significantly enhance both offensive and defensive cyber capabilities. This development echoes concerns about Anthropic’s Mythos, which had demonstrated the ability to breach secure National Security Agency networks in certain scenarios.

Anthropic’s two advanced models, Mythos 5 and Fable 5, are currently restricted from public access due to national security concerns expressed by the White House. These fears are palpable across Washington, yet the strategic advantage lies with the United States.

The Asymmetry Favors America

Anne Neuberger, a former White House deputy national security adviser for cyber and emerging technologies, emphasized that AI is a dual-use technology, benefiting both sides: offense and defense. For the U.S., this means enhanced military capabilities, as well as advantages in intelligence collection.

Neuberger notes an asymmetrical advantage, with U.S. models offering a strategic edge if used effectively for both offense and defense.

She underlines the importance of first-hand visibility into AI advances, which reduces the likelihood of being surprised by adversaries’ developments. This perspective suggests a strategic advantage remains that favors U.S. interests.

Project Glasswing and Vulnerability Detection

Anthropic’s Project Glasswing uses Mythos to identify software vulnerabilities proactively. This initiative collaborates with partners like Amazon, Apple, Google, and NVIDIA. It has identified nearly 3,900 critical vulnerabilities in open-source projects, demonstrating its practical application in cyber defense.

Despite skepticism from some, such as security writer Bruce Schneier, these findings support the strategic logic of using advanced models to preemptively address vulnerabilities.

The Offensive and Defensive Dynamics

The red-team exercises involving Mythos were designed to test U.S. defense systems, uncovering weaknesses before an adversary could exploit them. Mythos successfully simulated cyberattacks, but the lack of active defenses in these simulations changes the interpretation from threat to opportunity.

Neuberger stresses that lag in deploying tools for defense remains a concern, particularly since U.S. infrastructure is mostly privately owned. This ownership complicates rapid AI deployment for cyber defense.

China’s Competitive Position

While the U.S. holds the frontier edge, China is progressing rapidly in the broader market with models that are adaptable and economically accessible. Neuberger highlights concerns over Chinese open-source models due to potential security and propaganda risks.

Michael Horowitz of the Council on Foreign Relations points out the rapid pace of AI adoption in China, signifying a significant challenge despite the U.S.’s current lead.

Ensuring AI Leads to Strategic Gains

Time is critical, with the U.K. AI Security Institute noting advancements in models like OpenAI’s GPT-5.5 equaling Mythos in certain simulation tests. Anthropic anticipates rivals fielding similar models soon, with potential for no safeguards.

In response, Neuberger emphasizes disciplined use of AI technology, focusing on classified testing, rapid vulnerability disclosures, and robust audit rules to harness AI’s advantages without compromising security.

The Five Eyes alliance’s warning underscores the urgency for defenders to leverage AI effectively. As AI continues to evolve, the nation that iterates and adapts swiftly will maintain a strategic advantage. The risk of complacency remains; ensuring private sector alignment with national security objectives is crucial.