Investigation finds more than 110 deaths within range of Anduril border towers

Why it matters

Anduril's border program exposes the accountability gap in public-sector AI: accurate classification has limited value when agencies do not measure whether alerts produce timely action. Contracts should track response outcomes alongside detections and coverage.

A MIT Technology Review investigation found that more than 110 people died within the advertised range of Anduril's autonomous border surveillance towers between 2021 and early 2026. The findings put a hard outcome against founders Palmer Luckey and Brian Schimpf's core product thesis: software should help government operators find people and act faster.

Luckey built Oculus around cheaper, rapidly iterated consumer hardware before Facebook acquired the virtual-reality startup in 2014. Schimpf previously built Palantir's Foundry product and led an autonomous-vehicle research program at Cornell University. They co-founded Anduril in 2017 with Trae Stephens, Matt Grimm and Joe Chen, applying Silicon Valley's software and hardware development model to slow government procurement cycles.

Border surveillance was an early proof point. Anduril's first hardware product paired cameras and radar with computer vision designed to classify and track people, vehicles and animals. Its Lattice software could generate alerts without requiring agents to continuously scan banks of video screens. Anduril moved from a pilot to a formal Customs and Border Protection program of record in 2020, an unusually fast trip through federal procurement.

The new investigation shows why deployment speed and alert volume are incomplete measures of whether that system works.

A death 360 feet from a tower

Jose Morales Bernal crossed into southern New Mexico on April 8th, 2024, one day before his 32nd birthday. He entered an area covered by three recently installed Anduril towers near Sunland Park, according to the investigation by MIT Technology Review and the Times of San Diego.

Employees at a nearby landfill saw Morales that morning. At 1 p.m., they saw him lying in the sand. At 4 p.m., after he had remained motionless, the workers called Border Patrol. Agents arrived 45 minutes later and found him dead. An autopsy attributed his death to environmental exposure.

Morales was found about 360 feet from the closest tower. Two more stood to the east and west. The investigation found that 18 other people died within range of those three towers during 2024, including five whose locations were visible from the tower nearest Morales.

The reporting does not establish whether that tower detected Morales, correctly classified him, generated an alert or delivered an alert to an available agent. Tower footage and related data are generally overwritten after 30 days unless preserved for an active investigation, according to the report. Morales died more than two years before the investigation was published on September 21st, 2026.

That uncertainty protects neither Anduril nor CBP from the larger finding. The system was installed to create awareness and prompt action in precisely that corridor. Landfill workers, rather than the network of sensors, software, control rooms and agent smartphones, initiated the response.

Anduril told the reporters that CBP operates the towers after delivery. Anduril also said proximity to a tower does not prove a missed detection because actual coverage depends on terrain, physical obstructions and virtual boundaries set by CBP. Those qualifications matter. They also underline how much of the advertised outcome depends on placement, configuration, staffing and field response after the hardware is installed.

Automation moved the bottleneck

Anduril's pitch addressed a real operational problem. Earlier tower systems forced agents to monitor dozens of screens while handling radio traffic, record checks and dispatch work. Agents interviewed for the investigation described broken cameras, information overload and control-room assignments that were routinely given lower priority than field work.

The autonomous tower was supposed to reduce that burden. A 33-foot solar-powered Sentry tower uses radar to identify movement, points cameras toward the activity and applies computer vision to classify what it sees. Alerts can then appear on government-issued smartphones, allowing nearby agents to respond.

Anduril said in September 2024, when it announced the deployment of its 300th CBP tower, that the systems covered about 30% of the southern land border. Anduril also claimed its towers had autonomously identified hundreds of thousands of crossings and contributed to humanitarian responses. Those are Anduril's figures, and they do not reveal how often a detection led to no response, a delayed response or no documented outcome.

Former agents described gaps on both sides of the interface. Smugglers learned to cover groups with tarps, blankets or netting, causing Anduril's software to classify moving rafts or groups as unidentified objects that did not trigger audible alerts. In New Mexico, early misses led some agents to distrust the automation and operate cameras manually, increasing the risk of further missed alerts.

Detection could also work as designed without producing an immediate interception. CBP decides which areas towers monitor, how alerts are prioritized and whether agents should move toward a group. Agents told reporters they might wait to see which route people took, conserve staff for a larger group or plan an apprehension at a more accessible location.

The product boundary is clear. Anduril can supply a detection and tracking system. CBP controls the operational chain that turns an alert into a rescue, apprehension or unanswered notification. Evaluating only model accuracy leaves the decisive part of the system unmeasured.

Billions more are moving toward the same model

The investigation cross-referenced nearly 4,000 locations where human remains were found with almost 600 surveillance towers. It identified more than 1,050 deaths within tower range between 2015 and early 2026, spanning systems supplied by Anduril, Elbit Systems and General Dynamics. Deaths occurred within the advertised range of nearly two-thirds of the towers analyzed.

Terrain explains part of the gap. The reporters' topographical analysis found that some towers could see as little as 10% of their advertised coverage area. Most of the recorded deaths, however, were not in terrain blind spots.

A 2026 Government Accountability Office assessment counted 803 fielded towers across CBP's integrated surveillance program. CBP plans to expand toward roughly 2,300 towers through new long-range systems and upgrades that add autonomous functions to older equipment. MIT Technology Review reported that the government estimated in 2023 that 803 towers would cost $6.2 billion over their lifetimes. CBP plans to spend another $1 billion on 1,497 towers by 2034.

Anduril has strong incentives to keep autonomy at the center of that expansion. Its border tower contract is worth up to $1.1 billion, according to the investigation. In May, Anduril closed a $5 billion Series H at a $61 billion valuation, led by Thrive Capital and Andreessen Horowitz. Schimpf said Anduril generated about $2.2 billion in revenue during 2025, roughly twice its 2024 figure.

The border business also helped establish Anduril's broader argument that a venture-backed contractor could ship government hardware quickly and turn autonomy into a repeatable platform. That argument has since carried Anduril into larger defense programs. RuntimeWire reported in June that Anduril won an Air Force production slot for its FQ-44 autonomous fighter, moving Anduril deeper into programs once dominated by traditional contractors.

The surveillance investigation does not show that Anduril's software failed in every death near a tower. It shows that the federal government and its contractors have promoted detections, coverage and deployments without a comparable public accounting of what happened after an alert should have been possible.

Luckey and Schimpf built Anduril around the belief that better software can make government systems faster and more capable. The border program now supplies a demanding test of that belief. A surveillance network succeeds when the person it identifies is reached in time, and the available measurements rarely follow the system that far.