AI Moves Into the Physical World, the Military Watches Closely
Anthropic’s Claude programs a robodog 20 times faster than its fastest human team; six companies build an autonomous drone-killing ground robot in 28 hours at the US Army’s Operation Jailbreak. Also: Ukraine opens its combat AI data to 100 firms, Germany turns drones into standoff mine layers, and Ukraine’s drone model spreads to Africa.
Claude Programs the Robodog
Anthropic’s Claude Opus 4.7, operating without human assistance, completed a series of robotics programming tasks 20 times faster than the fastest human team in the equivalent experiment run less than a year ago. Project Fetch, Anthropic’s internal test using an off-the-shelf robotic quadruped, originally showed that Claude-assisted human teams outperformed unaided ones. The new result removes the human from most of the loop entirely: Claude connected to the robot’s sensors, wrote working control code on the first attempt, and produced roughly ten times fewer lines of code than the human teams while achieving equal or better results. It still struggled with the closed-loop physical task of actually guiding the robot to nudge a beach ball back to a target zone, a task that requires moment-to-moment sensory correction of the kind humans find intuitive. But Anthropic notes the trajectory: models go from helping humans, to being helped by humans, to doing things themselves. That pattern has already played out in cybersecurity. It is now beginning to take shape at the intersection of AI and the physical world.
A Drone Killer in 28 Hours
Six defence companies assembled an autonomous counter-uncrewed aerial system (UAS) ground robot in approximately 28 hours at Fort Carson, Colorado, during the US Army’s Operation Jailbreak, a sprint designed to get disparate military systems communicating on a shared network. AZAK provided the wheeled chassis, Allen Control Systems mounted its Bullfrog autonomous machine-gun turret on top, Havoc supplied the autonomous driving software, Leonardo DRS added radar, Picogrid handled system integration, and everything ran through Anduril’s Lattice battle management platform. The result is a hunter-killer pair: one robot senses the threat, the other shoots it down, both moving autonomously through contested terrain without requiring troops in static positions. The build was made possible because Jailbreak required all participating companies to share their interfaces openly, a direct departure from decades of siloed defence acquisition. The Army has not yet announced procurement intent, but the two lead companies have already signed a memorandum of understanding to scale joint production to the thousands if asked.
Ukraine Opens the Data
Ukraine’s Ministry of Defence has given more than 100 companies access to Brave1 Dataroom, a secure platform containing visual and thermal imagery of aerial targets collected under real combat conditions. The datasets cover different target types across varying weather, time of day and sensor configurations, with particular focus on detecting and neutralising Shahed-type strike drones. Dozens of AI and computer vision systems trained on the platform are already in operational use on the frontline, handling tasks from autonomous drone guidance to GPS-denied navigation under electronic warfare conditions. A companion platform, Avengers Labs, allows foreign developers to train models on millions of annotated frames from live missions. The practical effect is to turn two years of the world’s most intense drone war into a training corpus available to allied industry.
Germany’s Mine-Laying Drone
German firm Dynamit Nobel Defence unveiled Eggbox at Eurosatory, a modular pod that allows any heavy drone to carry and deploy AT2+ anti-tank mines from standoff distances of several dozen kilometres. The pod holds four mines, each weighing 2.22 kg, fitted with a magnetic fuze and an explosively formed penetrator warhead capable of penetrating approximately 140mm of armour. The same mine was transferred to Ukraine in 2022 and contributed to the destruction of 120 Russian tanks in three weeks in Kherson Oblast. Attaching it to a drone removes the need for ground forces or artillery to enter defended ground to lay obstacles. Ukraine is the obvious proving ground for a system explicitly designed around the kind of standoff mining that Ukrainian forces have been improvising with aerial drones for two years.
Ukraine’s Model Goes to Africa
Ukraine is actively developing its drone export market, with several African states looking to acquire systems built on two years of intensive battlefield adaptation. Ghana has expressed interest in Ukrainian surveillance drones for its Sahel borders, and President Zelensky has signalled that export permits for drones, missiles and shells are being expedited for African clients. The trajectory of that adaptation is documented: consumer Mavic quadcopters weaponised in 2022, first-person view drones at scale in 2023, fibre-optic systems from 2024. Analysts at the S. Rajaratnam School of International Studies note that Islamic State affiliates in Mali are already emulating FPV and fibre-optic tactics learned from the Russia-Ukraine conflict, while unconfirmed reports point to Ukrainian drone instructors operating in the DRC and South Sudan. What began as a European land war is functioning increasingly as a global technology distribution network.
Image – Anthropic
