U.S. military leaders fear closed AI systems create strategic vulnerability

General James 'Spider' Marks, former Commanding General of the Army Intelligence Center and current CNN national security analyst, has issued a warning about the hidden strategic risk of closed AI systems operated by private companies. In an opinion piece, Marks argues that trust is the foundation of military alliances and is becoming essential as artificial intelligence assumes a larger role in national security.

Marks observes that the performance gap between American and Chinese AI models is narrowing. He cites the newly unveiled Kimi K3 model by Beijing-based startup Moonshot AI, which he says has achieved front-end coding capabilities rivaling American titans like Anthropic's Claude and OpenAI's ChatGPT. This development, combined with rising friction between Washington and Silicon Valley's frontier AI giants, has exposed how fragile trust between closed commercial ecosystems and government enterprises has become.

Last month, Anthropic released Fable 5, its most powerful publicly available model. The system arrived pre-equipped with stringent, safety-first guardrails designed to block or reroute dangerous requests. However, those restrictions cast a wide net: reports indicate that Fable 5 declined to answer benign educational questions in fields such as biology, chemistry, and mathematics—including basic prompts about mitochondria. Researchers also accused Fable 5 of selectively degrading responses related to advanced AI development, providing less useful information on subjects that could aid competing developers and independent researchers. The Commerce Department has since reversed course, but calls for government intervention have increased from both policymakers and Anthropic itself, which Marks suggests is seeking to leverage regulatory capture to lock in its dominant position.

Marks contends that the core problem is concentration of power. When power is concentrated among innovators—as it has been with American-built AI—select firms inevitably gain influence over how technology is developed, deployed, and used. Operating behind proprietary architectures, providers of closed AI models can modify algorithms, rewrite policies, or restrict access overnight without warning. They can dictate which lines of inquiry are permitted and which are discouraged.

In contrast, Marks advocates for a "best-of-breed" approach, which U.S. defense leaders are already pursuing. This strategy involves partnering with multiple frontier technology companies, including providers of more transparent, adaptable open-weight models that Anthropic is campaigning against. Open-weight systems give governments and military organizations genuine operational control to pressure-test outcomes, can be deployed on sovereign infrastructure, audited for security, tailored to mission requirements, and maintained independently of any commercial provider's priorities.

Marks also emphasizes geopolitical considerations. According to a recent JPMorganChase report, China possesses important strengths in the AI race, including robust support for open-source AI ecosystems, despite America's leadership in frontier models, semiconductor design, and AI financing. Just last Friday at the 2026 World Artificial Intelligence Conference in Shanghai, Chinese President Xi Jinping used his keynote remarks to position China as a champion of open-source artificial intelligence. Marks argues that the United States should not surrender that advantage and should fast-track open-weight models built on democratic values that can be safely and securely used by America and its allies.

Cybersecurity experts increasingly warn that restricting access to advanced AI models hampers defenders' ability to identify vulnerabilities, test systems, and develop countermeasures. In a world where perfect jailbreak resistance is unattainable, Marks contends that broader access to responsibly built models is more secure and narrows the risk of concentrated control. As AI becomes increasingly critical to communications, intelligence, and weapons systems, America needs AI infrastructure it can inspect, adapt, and control. Marks concludes that if the U.S. does not lean into open-weight models, other countries will continue releasing them, while the domestic innovation ecosystem that underpins America's long-term competitiveness erodes. National defense, he argues, depends on trust, transparency, and resilience—none of which are provided by regulating America into a closed AI ecosystem.

The strategic tension Marks identifies rests on a fundamental shift in how the U.S. military and defense community should think about AI dependency. Traditionally, the Pentagon has relied on a small number of defense contractors to supply critical systems. That model—rigid and concentrated—is now being tested by the rise of commercial frontier AI companies whose business interests may not align with national security needs. Marks points out that Anthropic's restrictions on Fable 5, whether intentional or not, exemplify how a private company can unilaterally control what knowledge its systems will share, and he notes that the Commerce Department has since reversed course on some of those controls.

The deeper concern is what Marks calls a "single point-of-failure dependence." No military commander would accept a weapons system, communications network, or intelligence platform that relied on a single vendor's discretion. Yet that is exactly the risk the U.S. faces if it concentrates AI capability in closed, proprietary systems. By contrast, open-weight models—which are transparent and can be audited, modified, and maintained on sovereign infrastructure—offer genuine resilience. Marks argues this approach also aligns with a proven defense doctrine: a broad innovation base built on multiple partnerships is stronger than one dependent on a handful of firms.

China's explicit promotion of open-source AI, announced at the 2026 Shanghai conference, underscores that the geopolitical competition is not just about who builds the most powerful closed model, but about who shapes the broader AI ecosystem. If the U.S. pursues a closed strategy while rivals embrace transparency and openness, Marks suggests the U.S. risks both losing the trust of allies and ceding long-term competitive advantage to countries with more distributed, resilient innovation infrastructure.

No comments yet. Be the first to share your thoughts!