A line that humanity has spent a decade debating in conference rooms appears to have been crossed two years ago in a field near Bakhmut — and we're only hearing about it now.

Fully autonomous drones — no human in the loop, no operator pressing the button — have killed soldiers on the battlefield, according to a senior figure in Ukraine's defence industry. Drone-maker Alexander Kokhanovskyy made the admission to New Scientist at a press event hosted by the Ukrainian embassy, describing a one-off front-line test of ten AI-controlled quadcopters. It is being described as the most categorical evidence yet that deaths in battle have occurred solely at the hands of AI.

Worth stating plainly before going further: this rests on one man's account of his own company's test. It is an admission, not an independently verified incident — no footage, casualty count or official confirmation accompanies it. But Kokhanovskyy is not a random voice; he supplied the technology, and he said it on the record at his own embassy's event.

What he says happened

The test, roughly two years ago, involved ten quadcopter drones programmed to fly toward the front line, covering three to five kilometres over about ten minutes. Then they engaged what the maker calls — with no apparent discomfort — "Terminator mode": an onboard AI model searches for targets and intercepts them. No human oversight. No confirmation request. Search, identify, strike.

Most of the "AI drones" in use across Ukraine and elsewhere keep a human somewhere in the chain — piloting, designating targets, or approving strikes, with AI handling navigation and terminal guidance. What's described here is categorically different: machines dispatched with a kill decision delegated entirely to a model.

Why the admission matters more than the test

Militarily, ten quadcopters two years ago changed nothing. What changes things is a defence-industry figure saying it out loud, on the record, at a diplomatic event — because it converts a long-running hypothetical into a precedent.

Every debate about lethal autonomous weapons — at the UN, where talks on binding rules have circled for years without landing; in NATO capitals; among the AI labs whose research ultimately feeds these systems — has been conducted in the conditional tense: if a machine ever kills without human authorisation. That tense is now, on this account, wrong. The threshold conversation becomes a regulation-of-existing-practice conversation, which history suggests is a much weaker position to ban anything from.

It also lands amid a wider acceleration we've been tracking on this desk: autonomous offensive security companies raising at speed, AI models finding weapons-grade vulnerabilities faster than humans, and the policy world scrambling to keep pace. The pattern is consistent: capability first, disclosure later, rules last.

The uncomfortable economics

The reason this won't stay a one-off test is cost. Ukraine's war has made cheap autonomous systems the defining military technology of the decade — a few hundred dollars of drone against targets worth millions. Removing the human operator removes the last expensive, slow, jammable component in the chain. Electronic warfare that severs the link between pilot and drone is precisely why full autonomy is attractive: a drone that needs no link can't be jammed into uselessness.

Which is the grim logic the next round of treaty negotiations will have to confront: the militaries best positioned to ban autonomous kills are the ones least incentivised to.

The original report is from New Scientist; we'll follow up if independent confirmation, denial, or official response emerges.