American models, Chinese military training data
The day's dominant story came from Reuters, in an exclusive by Eduardo Baptista: Chinese military researchers have been using outputs from leading American AI systems — including models built by OpenAI and Anthropic — to help develop and train domestic AI for defense purposes. Reuters based the finding on its own review of the researchers' work.
The significance is less about any single leaked file than about the mechanism. Export controls were designed around hardware — chips you can count, crate, and stop at a border. Model outputs are none of those things. If a foreign defense lab can query a commercial American model and use what comes back as training material for its own systems, the frontier lab becomes an input to a supply chain nobody licensed. Expect that framing to show up quickly in Washington, where the chip-centric theory of AI containment has just been handed an awkward counterexample.
The timing sharpens it. Chinese President Xi Jinping used the same news cycle to call for greater military use of autonomous and artificial intelligence technologies, according to Chinese state media relayed by Bloomberg's Josh Xiao. Demand signal and supply story, landing on the same day.
The trade war grows legs
Washington's response to Chinese technology has moved past semiconductors and into machines that walk. Forbes reported on July 31 that the US trade fight has extended to humanoid robots and to equipment that keeps American data centers powered; a WAFB account framed the action as a ban on foreign-made humanoid robots. The two reports differ in how broadly they characterize the scope, and that ambiguity matters — a China-specific restriction and a blanket foreign-manufacture ban imply very different things for allied suppliers and for US firms sourcing abroad.
What's consistent is the direction of travel. Robotics and grid and cooling infrastructure are being treated as strategic categories rather than commercial ones. Data center power equipment is the quieter half of that pairing and possibly the more consequential: it is the physical bottleneck under every AI buildout in the country.
The scrutiny is reaching further down the stack than robots. CNBC reports that US lawmakers have asked DoorDash for information about its use of Chinese AI models, a request stemming from a joint investigation by a House committee. A food delivery company is now a subject of national security inquiry because of which models sit inside its software — a useful marker of how far the definition of "critical technology" has stretched.
Underneath all of it is the mood The Guardian described: China's rapid technological progress is scrambling assumptions in Silicon Valley boardrooms and in Washington policy offices alike. Today's stories are what that anxiety looks like when it becomes policy.
Water utilities hit, Iran suspected
Wired reported in its weekly security roundup that water systems across seven US states were struck by cyberattacks likely tied to Iran. The account is brief and details are thin — no confirmed attribution, no public accounting of operational impact. Water utilities remain among the softest targets in American critical infrastructure: thousands of small, underfunded operators running aging control systems. Seven states in one reporting window is a pattern, not an incident.
Pentagon buys drone killers, Army scales its network
On the procurement side, DefenseScoop reports the Pentagon's counter-drone task force has awarded a contract worth up to $500 million for SkyValor, a "detect and defeat" system built to spot hostile drones and bring them down. The award follows testing at the border — a reminder that counter-UAS is now a domestic mission as much as an expeditionary one.
Also via DefenseScoop: the Army is preparing to push parts of its Next Generation Command and Control effort, NGC2, out to more units. The internal verdict — "ready, but not done" — is candid. Not every piece of the ecosystem is guaranteed, and the service is scaling anyway.
The thread
Autonomy, attribution, and access all pulled in the same direction today. The US is trying to draw borders around a technology whose most valuable component — a model's output — crosses them by design.