Key Takeaways
- The clash between Anthropic and the Pentagon shows that private AI firms can effectively veto how the U.S. military uses advanced models.
- Anthropic’s “Mythos” model is reportedly capable of autonomously discovering and weaponizing cyber vulnerabilities, raising serious security concerns if left unchecked.
- The current procurement model—renting access to closed, proprietary systems—leaves the Department of Defense dependent on vendors’ governance, risk tolerances, and commercial incentives.
- Competitors such as China are fielding open‑source, state‑aligned AI systems that can be rapidly adapted for military use, creating an asymmetric advantage.
- To retain military dominance, the United States must develop controllable, high‑performance, open‑source AI models that the government and trusted allies can audit, modify, and deploy without external constraints.
- Ethical guardrails and oversight must remain, but decision‑making authority over AI use in warfare should reside with elected officials and military leaders accountable to the public, not corporate acceptable‑use policies.
The United States is moving into a new era of strategic competition where artificial intelligence is no longer a nascent technology but a core determinant of military power. In this AI arms race, speed, capability, and—most critically—control dictate who can prevail. The recent rupture between AI startup Anthropic and the Pentagon underscores why control matters: when the developer of the Claude family and its powerful successor, Mythos, sought to impose strict limits on how its technology could be employed, the Department of Defense pushed back, insisting on the right to use any lawful AI tool for national defense. The inability to reconcile these positions led to Anthropic being labeled a supply‑chain risk and the Pentagon forced to seek alternative sources.
Anthropic’sMotivation for restricting Mythos stemmed from fears that the model could autonomously identify and weaponize undiscovered cybersecurity vulnerabilities, effectively giving cybercriminals an open‑season toolkit if proper safeguards were absent. While the company has kept tight reins on access to Mythos, the episode exposed a broader structural weakness: the U.S. defense establishment purchases AI capabilities as a service but does not own or govern the underlying models. Training, testing, and continual improvement remain locked inside private firms that operate under their own governance frameworks, risk appetites, and profit motives. Consequently, a handful of unaccountable companies wield de‑facto veto power over how the United States can deploy one of the most consequential technologies of our time.
This dependence creates a dangerous misalignment between the rapid tempo of modern warfare and the sluggish, approval‑heavy processes imposed by external vendors. In a conflict where iteration cycles are measured in weeks rather than years, any delay caused by corporate policy shifts, licensing renegotiations, or the threat of sudden contract termination can be exploited by adversaries. China and its partners exemplify the alternative approach: they are aggressively fielding open‑source models—such as the DeepSeek family—that can be freely modified, extended, and integrated across military, intelligence, and allied networks. Because these systems are not shackled by corporate acceptable‑use policies, they can be tailored to operational needs at speed, giving China an asymmetric advantage while the United States remains bogged down in contractual negotiations.
The solution is not to abandon the private sector, which remains a vital engine of AI innovation, nor to discard essential ethical considerations regarding autonomy, targeting, surveillance, and escalation. Rather, the United States must recognize that renting closed, proprietary AI is insufficient for the demands of strategic competition. Washington should invest in the creation of high‑performing, secure, and adaptable open‑source AI models that the federal government and its closest allies can fully control, audit, and deploy without external constraints. Such models could be developed through government‑led initiatives, partnerships with trusted research institutions, or alliances that produce interoperable, defense‑specific open‑weight models. Procurement strategies would need to prioritize transparency, modifiability, and long‑term stewardship over short‑term convenience.
Importantly, establishing these capabilities does not eliminate the need for rigorous guardrails. Debates about the appropriate role of AI in warfare—ranging from lethal autonomy to escalation risks—must continue, but they should be led by elected officials and military leaders accountable to the American people, not dictated by the terms of service of private vendors. Building effective public‑private partnerships will require trust, aligned incentives, and clear processes that share responsibility for both the benefits and the risks of AI deployment.
If the United States fails to act, the Anthropic episode may prove not an isolated incident but a harbinger of a widening capability gap. By securing sovereign control over its AI foundation—just as it does for ships, weapons, and communications—the nation can preserve its capacity to innovate rapidly, respond decisively to threats, and maintain military dominance in an era where artificial intelligence decides the outcome of strategic competition. The time to build that foundation is now.