AI border towers logged 1,050 deaths — CBP never audited them
MIT Technology Review published the first comprehensive map of deaths along the US border's AI surveillance towers today. The finding is not that the towers are imperfect — it is that nobody was ever required to check.
MIT Technology Review mapped more than 1,050 deaths near the government's "virtual wall" of AI-enabled surveillance towers, and found that US Customs and Border Protection has never audited whether any of those towers should have detected the people who died. The clearest case in the investigation is José Morales Bernal, who crossed into southern New Mexico in April 2024 and died 360 feet from the closest of three Anduril towers watching that stretch of desert. Landfill workers spotted him; agents arrived 45 minutes after they called, and an autopsy recorded death by environmental exposure. Eighteen other people died that year inside the range of those same three towers, five of them visible from the nearest one.
The numbers scale from there. MIT's analysis puts somewhere between 500 and 700 deaths since 2015 within range of the older General Dynamics RVSS cameras, and roughly 350 near Elbit's integrated fixed towers in Arizona — a tally the reporters describe as an undercount, because CBP publishes no map of its own network and the agency's internal death-tracking system has been flagged by the GAO for missing cases since 2022. Agents told MIT the Anduril towers initially missed 10% to 15% of what they should have caught, and that smugglers defeated the classifier by covering rafts and groups with tarps and netting — the system logged motion as an "unidentified object" and stayed silent, which is exactly the alert agents were relying on it to raise. CBP deletes tower footage after 30 days unless an internal review is opened, and there is no process anywhere that checks a migrant death against tower positions. CBP assistant commissioner Hilton Beckham said autonomous towers "complement physical barriers" and that "the technology does not make law enforcement decisions," while Representative Delia Ramirez called for terminating the tower program outright. The spending is going the other way: CBP plans to spend $1 billion on 1,497 more towers by 2034, with border encounters down to 128,009 in the first eight months of 2026 from more than 2.4 million in the same period of 2024. We argued the same failure mode from the other end last week — Deep Dive — The human in the loop was the weakest link — detection without a response is not a safety system.
SenseTime shipped the production version of SenseNova U1 Pro, its flagship image-creation model, with output up to 8K resolution and support for unusual aspect ratios. The pitch is design work rather than pretty pictures: the model reasons through an interleaved text-image chain of thought, and SenseTime claims the results hold up on text rendering and layout next to the top overseas models. The company's own testers ran it against real deliverables — a music-festival poster, a character sheet for an animated series, a designer-toy product shot, a set of recruiting materials — and reported the harder part working: keeping one visual style across posters, standees and brochures while rearranging information for each format. Local edits held up better than expected too, including swapping a subject's clothing and background while preserving pose, and changing individual characters inside a finished poster. It is live in SenseTime's Raccoon office agent and its SenseNova API. We covered the lighter open-weight sibling in August — SenseTime open-sources SenseNova U1.5-Lite multimodal model.
Huawei Cloud upgraded CodeArts for HarmonyOS developers, adding a HarmonyOS-specific coding model, an end-to-end coding agent, and a developer practice center — a stack the company calls the first AI coding agent built exclusively for the HarmonyOS ecosystem. Huawei says the model was trained on more than a million HarmonyOS source assets and tuned for ArkTS and HarmonyOS UI conventions, and claims it cuts error rates per thousand lines of code by more than 80%. The agent covers requirement analysis through compilation, and carries built-in HarmonyOS documentation, third-party library conversion and mini-program-to-meta-service skills. The addressable audience is real: Huawei says HarmonyOS has passed 11 million registered developers, and a vertical model for one framework is a narrower but sharper bet than another general coding assistant. We covered Huawei's earlier HarmonyOS agent play — Huawei's open-source WorkSwarm brings swarm agents to HarmonyOS PC.
What to watch: whether the border-tower investigation produces an audit demand with a bill attached, and whether a framework-specific coding model outperforms general models in practice rather than on Huawei's own benchmark.
Should a government surveillance program be required to audit itself against the outcomes it promised? Tell us in the comments.
