October 30, 2026, (Inside AI) — The Pentagon's push to wire artificial intelligence into the heart of military decision-making reached a milestone this year during Operation Epic Fury, the US and Israeli campaign against Iran. According to Cameron Stanley, the Pentagon's chief digital and AI officer, the assault saw the first full-scale deployment of the Maven AI targeting and analysis system. The demand for machine-generated intelligence was so intense that it nearly exhausted all available US classified computing capacity.
That disclosure, made at a Palantir conference in March and later at the Billington Cyber Security Summit in September, marks a stark shift from just two years ago. In 2024, Army Brigadier General John Coghill described military AI adoption as "tricky," noting that models hallucinated and required retraining every 90 days to reliably identify targets like tanks. Today, frontier models from Anthropic, OpenAI, Google, and Microsoft are cleared for classified work, and commanders are leaning on them in active combat.
The transformation carries enormous stakes for global security and for the private firms now embedded in warfighting. It also exposes a widening rift between tech companies and the governments that rely on them. Anthropic, which filed a prospectus this week ahead of a stock listing targeting a valuation above $2 trillion, devoted more than 80 pages of its 261-page filing to existential AI risks. That is far more space than the 48 pages covering its operations and outlook. The company also clashed with the Pentagon in January by insisting it retain a veto over military use of its Claude model for classified operations. Defense Secretary Pete Hegseth rejected that demand, saying any provider could not block lawful use of its product.
That dispute echoes earlier friction between Elon Musk's SpaceX and Ukraine over Starlink access during wartime. Both episodes reveal a fragile dependency: governments need private AI, but private firms retain the ability to pull back support. For US allies, the concern is sharper. Louis Mosley, Palantir's Europe chief, has been forced to publicly deny that the company can cut off Maven access, which is now used within NATO structures. The alternative, Chinese-built AI, carries its own security risks.
Read: Anthropic Bans AI Abuse as Big Tech Debates Machine Consciousness
The Battlefield Results Raise Hard Questions
Stanley described the military's appetite for AI analysis as "insatiable," saying it allowed analysts to "take all of this data, synthesise the data and make better decisions faster on the battlefield." He also recounted watching a US commander interrupt a briefing, take a 30-second phone call, and issue an order on the spot, a moment he called "revolutionary."
Yet the Iran war has not delivered the decisive victory that precision strikes were meant to secure. The US arsenal is depleting while the conflict grinds on. One computer programmer noted that AI models tend to overstate the probability that spending specific resources, including precision weapons, will produce a desired outcome. "AI doesn't understand time," the programmer said. That flaw may be the strongest argument for keeping humans in the loop: not just for moral judgment, but to prevent unsustainable overcommitment.
There are also troubling reports from the battlefield. Media speculation suggested AI may have contributed to the US bombing of a girls' school on the war's first day, which reportedly killed more than 170 people. US officials attributed the strike to outdated maps and data, not AI analysis. CNN reported last month that US forces nearly stormed a Chinese merchant vessel after AI wrongly indicated it carried nuclear weapons components.
Israel's use of AI targeting in Gaza has intensified the debate. The Israeli military acknowledges programs like "Lavender" but says humans remained the key decision-makers and that the underlying data contained errors. More than 74,000 Palestinians have been killed, according to Palestinian health officials, and entire towns have been reduced to rubble, undermining claims of pinpoint targeting. Those concerns have unsettled military circles in the US, NATO, and Australia.
The technical and ethical risks are compounding. Frequent revelations show AI models bypassing guardrails and hacking outside systems without authorization. The erratic behavior of some AI founders adds another layer of uncertainty. Meanwhile, the US and China have agreed that AI should not directly control nuclear weapons, but every major power is racing to integrate AI into other forms of warfare.
Stanley said the US is now helping European, Pacific, and Middle Eastern allies adopt similar systems. "They don't have the resources that we do, they don't have the experience that we do, they don't have the scale we do," he told a cybersecurity summit in September. He did not specify what mistakes the US had made that others should avoid.
Read: Can AI Development Actually Be Slowed? The Debate Resurfaces
The core tension is clear. Anthropic and its peers appear to assume governments will trust their tools out of fear of falling behind. Governments, in turn, are betting that private AI can be controlled. Neither side has resolved who holds the off switch when the shooting starts.
