Between 2015 and early 2026, more than 1,050 people died within range of border surveillance towers that were supposed to catch them before they died. The towers did not. They were online, pointed at the right stretch of desert, and operated by companies whose marketing copy promises round-the-clock, AI-augmented awareness of who is moving through the brush. The deaths kept happening anyway. That is the core finding of a 15-month investigation by MIT Technology Review called “Dying on Camera,” and it deserves a closer look than a roundup mention can give.
What the “virtual wall” actually is
The shorthand is a network of fixed surveillance towers along the US southern border, most of them run by US Customs and Border Protection. Three vendor stacks sit inside that network. Anduril sells autonomous “Sentry” towers, which the company says use AI to detect and classify people, vehicles, and animals, then push alerts to Border Patrol agents. General Dynamics operates the Remote Video Surveillance System (RVSS), a network of camera towers that CBP has been buying and upgrading since the early 2010s. Elbit Systems of America builds the Integrated Fixed Towers (IFT) program. Together, those three systems account for the bulk of the approximately 800 towers the federal government has deployed, according to the investigation.
The price tag is enormous. The government’s own estimate puts the lifecycle cost of existing tower plans at about $6.2 billion. CBP’s current plan calls for 1,497 more towers by 2034, at a price tag of roughly $1 billion. A July 2025 spending bill added $2.7 billion for border security tech. Anduril’s contract ceiling has climbed to about $1.1 billion. RVSS, the longest-running program, was projected to cost $3.7 billion back in 2020, on top of a $429 million initial rollout and a $103 million 2013 contract ceiling for General Dynamics that has since been expanded. Elbit’s IFT contract started at about $145 million.
That is what a “virtual wall” costs: several billion dollars of camera, sensor, and AI-classification infrastructure, paid for by Congress and signed off on by three different administrations.
How MIT Tech Review matched deaths to towers
The investigation’s database is the first public cross-reference of migrant remains locations against tower locations. The methodology piece lays it out in detail.
The death side drew on public records requests to 17 Texas county sheriffs (14 produced usable records), the Pima County Office of the Medical Examiner, No More Deaths records for parts of California, Arizona, New Mexico, and Texas, and Humane Borders’ Arizona dataset. The total came to nearly 4,000 cases.
The tower side came from the Electronic Frontier Foundation, which has tracked border surveillance towers since 2023 and published a database in April 2026. EFF’s director of investigations, Dave Maass, contributed tower mapping and satellite imagery to the project. MIT Tech Review then added removed towers, ground-truthed the rest with Google Street View and EFF photos, and ran a topographical “viewshed” analysis to flag terrain that physically blocks a tower’s line of sight.
The match rule was simple and conservative: a death counted if the remains were within the tower’s published surveillance range and the tower was present during the estimated window of death. Roughly 600 towers were analyzed, against roughly 4,000 remains locations. Anthropic’s Claude, via API, was used to extract coordinates from sheriff and medical-examiner PDFs; outputs were then hand-verified. The 15-month investigation ran from mid-2025 through publication on September 21, 2026, in collaboration with Times of San Diego.
What the towers missed
The headline number is the 1,050+ deaths in range. The breakdown by vendor is sharper. Since 2015, between 500 and 700 people died within range of RVSS towers; nearly 350 died within range of Elbit IFTs; more than 110 died within range of Anduril’s autonomous towers since 2021. Of the 53 IFTs the team analyzed, 52 had at least one death within range.
The investigation documents four interlocking failure modes:
Broken equipment. CBP told Congress that 66% of first-generation towers were unusable at some point in 2024. By January 2026, 30% remained broken. A former agent quoted in the piece estimated that 10% to 15% of apprehensions were initially missed by the cameras.
Failed detection algorithms. In one incident, a group was misclassified as cattle. In another, a car crash on the Rio Grande near a tower never registered as an event.
Topographical blind spots. The viewshed analysis found that some Anduril towers in the Otay Mountain Wilderness could see just 15% to 25% of their advertised coverage area. In El Paso sector sectors, visibility ranged from roughly 80% to 90%.
Unresponsive agents. Former Border Patrol chief Jason Owens, who led the agency from June 2023 to March 2025, told the reporters the system never reached “set it and forget it.” Other current and former agents described units that “hang back” rather than intercept.
The named cases make the numbers concrete. José Morales Bernal died of environmental exposure on April 8, 2024, in the New Mexico desert near Sunland Park, the investigation reports. Three Anduril towers were within roughly 1.3 miles. The closest was 360 feet away. He was found by landfill workers, not by agents. Eighteen other people died in range of those same three towers in the same year.
In 2022, two-year-old Eloah Lenda and her father Daniel drowned in the Rio Grande near Del Rio, Texas. One camera system was about 100 feet from where Daniel fell in, with two others about a third of a mile away.
What This Means
The findings expose a contradiction at the heart of the program. The vendors sell detection; the operators buy coverage; the public is told the system “watches” the border. MIT’s data says the watching does not catch people who are dying in the cameras’ advertised fields of view. CBP assistant commissioner for public affairs Hilton Beckham defended the program to the reporters, saying autonomous towers “use artificial intelligence to detect and classify people, vehicles, and animals and alert Border Patrol agents.” The investigation’s reply is implicit in the 1,050 figure: the alerts, when they fire, are not converting into interventions.
That contradiction has a name in the investigation: a surveillance failure. The companion policy-recommendations piece lays out four asks: conduct a comprehensive audit of deaths near the virtual wall, fix the way the agency tracks migrant deaths and remains, track cases when surveillance technology leads to an apprehension, and investigate deaths as potential surveillance failures. The recommendations echo what a 2024 DHS Office of Inspector General report had already flagged, and what a Government Accountability Office finding described as “inaccurate and unreliable” logging.
For a privacy-focused site, the deeper question is what the surveillance stack is actually for. If the system is meant to save lives, it is failing at scale. If it is meant to enable rapid response, it is failing on its own terms. If it is meant to provide general situational awareness, the price tag is hard to justify against the documented outcome. The investigation does not call the program a sham, and the reporters do not editorialise beyond the data. They do not need to. The data already says: a system that watches 1,050 people die and cannot catch any of them is not what the public was sold. It is the latest in a string of AI-surveillance rollouts whose performance falls short of the marketing - from commercial doorbell cameras that quietly feed police networks to behavior-recognition deployments that an adversarial-pattern shirt can defeat.
The Bottom Line
The MIT Technology Review investigation has produced the first public, sourced cross-reference of border deaths against border surveillance towers. The result is more than 1,050 deaths in range of towers built, sold, and procured by three of the largest defense and AI companies in the United States. The system failed not because it was not deployed, not because it was not pointed at the right terrain, and not because the cameras were off. It failed because the entire pipeline from camera to agent to intervention breaks down at multiple points. That is a finding worth sitting with, and worth pressing your representatives on the next time a border-tech contract comes up for renewal.