U.S. Special Operations Command chief Adm. Frank Bradley recently warned that troops must be careful about how artificial intelligence enters the delivery of military force. The harder question is no longer whether the Pentagon will use commercial AI, but who will control the software once it becomes part of a mission.
The answer should be settled before deployment: For mission-critical AI, the United States government must retain the contractual and technical authority to stop the system, freeze an approved version, return to a previous version, and continue essential operations if the vendor relationship fails.
Recommended Stories
This is becoming urgent. In 2025, the Pentagon’s Chief Digital and Artificial Intelligence Office awarded contracts to Anthropic, Google, OpenAI, and xAI, each with a $200 million ceiling, to develop advanced AI workflows for national security missions. The Pentagon has since expanded its use of commercial AI in classified environments. Private firms will supply much of the talent and technology behind an AI-enabled force. That is a strength, but only if dependence does not become delegated command authority.
The Starlink controversy in Ukraine showed the underlying problem. In 2022, Ukraine asked SpaceX to activate coverage near Sevastopol for an attack on Russia’s Black Sea Fleet. Elon Musk declined, citing the risk of major escalation. The initial public account that he had switched off existing coverage was later corrected: The disputed area had not been activated. The important fact is narrower. A private provider’s decision about where a service would operate directly affected a wartime option.
That does not prove SpaceX acted wrongly. Nor does it mean the Pentagon should seize private intellectual property or build every model itself. It means military leaders should never discover during a crisis that a decisive capability depends on a provider’s unilateral policy judgment, a remotely pushed update, or a service decision that was never resolved in the contract.
Every contract for mission-critical AI should therefore contain six rights. The government should control an emergency stop for systems it operates. It should be able to freeze the last approved model version and roll back after a defective update. Core functions should retain an offline or degraded mode. Vendors should give advance notice of material changes to models, safeguards, or access terms. Operational continuity should survive a contract dispute, corporate acquisition, bankruptcy, or expiration. And for the most critical functions, the Pentagon should hold a protected executable package or technical escrow sufficient to recover the approved capability without demanding unrestricted access to the company’s source code.
These protections should run in both directions. Vendors should not be forced to support unlawful activity, and the Pentagon should not receive a blank check to repurpose software beyond agreed terms. The contract must define lawful uses, prohibited uses, update authority, and emergency control before the system enters the operational chain. Clear rules protect companies from improvised demands while protecting commanders from surprise vetoes.
This requires no new AI regulator. Congress should direct the Pentagon to make these clauses mandatory in major military AI acquisitions and require a public explanation when an acquisition official grants an exception.
WILLIE NELSON IS WRONG. HE’S DECLARING WAR ON THE DATA CENTERS STREAMING HIS MUSIC
Software sovereignty does not require the government to own every line of code. It requires the government to retain four powers over any capability on which a mission depends: continue, stop, freeze, and recover.
The Pentagon is right to buy the best American technology rather than trying to recreate Silicon Valley inside the government. But procurement must preserve command. When commercial AI becomes military infrastructure, the off switch cannot belong to someone outside the chain of command.
Burak Oktenli is an independent researcher in Washington, DC. He holds a bachelor’s degree in computer science and engineering from the University of South Florida and an MBA, and he is pursuing a Master of Professional Studies in Applied Intelligence at Georgetown University. His writing has appeared in the Washington Examiner, the Modern War Institute at West Point, RUSI, and RealClearDefense.
