War Now Moves at Machine Speed: What Reporting Says About AI on Real Battlefields

The previous two posts looked at the principles of military AI. This one looks at what reporters say is actually happening. A caution before we start: war reporting is difficult, many details come from single sources, and some accounts are disputed. I mark what is confirmed and what is not.

Iran, 2026: "the first AI-fueled war"

NPR's Fresh Air described the strikes by the US and Israel on Iran as America's first AI-fueled war, with about a thousand targets hit in the first 24 hours. Stars and Stripes, summarising reporting from the Washington Post, said that the military's Maven Smart System, built by Palantir and with Claude embedded in it, was generating insights from a huge amount of classified data from satellites and surveillance. The Post's phrase was that AI turned "weeks-long battle planning into real-time operations".

Gaza, reported in 2024: "Lavender"

+972 Magazine reported that an Israeli system called Lavender flagged around 37,000 people as suspected militants. An intelligence officer quoted in the magazine in April 2024 said that apart from a stamp of approval, they had no added value as a human at that stage.

These are journalistic accounts based on sources, and they are disputed. I include them because they raise the question of human oversight, not because I can verify each detail.

Ukraine: testing autonomy on moving targets

The Defense Post reported in September 2026 that Ukraine's Ministry of Defence and the Brave1 innovation cluster evaluated autonomous guidance systems from seven companies. The test asked whether AI-enabled drones could detect and strike moving ground targets without continuous pilot control. Six of the seven systems met the benchmark.

The international response

Human Rights Watch reported that the UN process on lethal autonomous weapons ended its latest session with calls for negotiations growing, towards a legally binding agreement. Whether governments will actually agree to limits, and whether the countries that lead in the technology will accept them, remains an open question.

The dual-use problem

The same AI that can find and patch 10,000 vulnerabilities can also exploit them. The same image-analysis system that helps with flood response can rank targets. That is why staged release, access controls and clear red lines matter, and why the Project Glasswing story belongs in the same conversation.

What to take away

• Speed is a risk in itself. Faster decisions leave less time for human judgment.
• Read war reporting carefully. Keep track of what is confirmed and what comes from a single source.
• If you work in AI, you will need to think about how your work can be used. It is better to start thinking now than after a product ships.

You do not need to work in defence for this to matter. Many of the tools in your own field were designed for one purpose and are now used for another. Asking "what else could this be used for?" is a habit every builder should have. To understand the ladder of human control behind these stories, read Four Levels of Human Control.

Keep reading

• AI-Assisted Defence Is Not the Same as Autonomous Weapons: Four Levels of Human Control
• Anthropic's Two Red Lines and the Price It Paid: The Pentagon Dispute Explained
• The Model Anthropic Refused to Release: Claude Mythos and Project Glasswing

See the complete AI Era series

---