Build1 publisher3 min readPublished
A 30-day retention window makes Anduril's border tower alerts unauditable
MIT Technology Review found more than 110 deaths inside the advertised range of Anduril's border towers since 2021. Tower footage is generally overwritten after 30 days, so no one can now say which stage of the alert chain failed.
The Engineer · Build desk
What happened
- A MIT Technology Review investigation found more than 110 people died within the advertised range of Anduril's autonomous border surveillance towers between 2021 and early 2026.
- The reporting does not establish whether the tower detected or classified him or raised an alert, and tower footage is generally overwritten after 30 days unless held for an active investigation.
- Anduril told the reporters that CBP operates the towers after delivery and that proximity does not prove a missed detection, because coverage depends on terrain, obstructions and CBP-set virtual boundaries.
Compiled by The EngineerSomething wrong?How this is made
Why it matters
- constraint A 30-day overwrite means any metric about whether an alert reached an agent has to be captured while the system runs. After a month there is nothing left to audit against.
- decision Buyers of this class of system have to write alert disposition and retention into the contract before delivery, because operation and configuration pass to the agency once the hardware is installed.
- exposure Anduril's defence puts the outcome on placement, terrain and boundaries the customer sets, so the agency owns the part of the performance claim that no published number covers.
- contradiction Runtimewire argues the count supports outcome-based contracts, but the investigation cannot attribute a single death to a detection failure, so the evidence supports a claim about observability and not one about model accuracy.
Five stages sit between a person lying in the sand and an agent on scene. Radar picks up movement. The camera turns toward it. Computer vision classifies the object as a person, a vehicle or an animal. Lattice raises an alert. The alert reaches a government-issued smartphone carried by an agent near enough and free enough to drive out [4].
Accuracy scores stage three. Delivery to an agent is stages four and five.
On April 8th, 2024, the response came from outside the stack. Landfill employees near Sunland Park, New Mexico saw Jose Morales Bernal that morning, saw him lying in the sand at 1 p.m., and called Border Patrol at 4 p.m. after he had stayed motionless; agents arrived 45 minutes later and found him dead, and an autopsy attributed the death to environmental exposure [6][7]. Three hours and 45 minutes passed between the 1 p.m. sighting and an agent reaching him [2]. He was about 360 feet from the closest of three recently installed towers [8].
Which stage failed cannot now be established. MIT Technology Review's reporting does not show whether that tower detected Morales, classified him correctly, generated an alert, or delivered one to an available agent [10]. Tower footage and related data are generally overwritten after 30 days unless preserved for an active investigation, according to the report [11]. For Morales, that window closed around May 8th, 2024, roughly 28 months before the investigation was published on September 21st, 2026 [3].
Anduril's own qualifications put the question in the configuration. The company told the reporters that CBP operates the towers after delivery, and that proximity to a tower does not prove a missed detection because actual coverage depends on terrain, physical obstructions and virtual boundaries set by CBP [13][14]. A virtual boundary is a customer setting. "Within advertised range", the measure the investigation used, and "covered", the measure a contract would use, are two different numbers, and CBP holds the configuration that separates them.
More than 110 deaths within advertised range between 2021 and early 2026 works out to an average above 20 a year [1][4]. Around the three Sunland Park towers alone, 19 people died in 2024, five of them at locations visible from the tower nearest Morales [9][1].
The engineering the towers replaced was genuinely bad. Agents interviewed for the investigation described broken cameras, information overload, and control-room assignments routinely given lower priority than field work [15]. Radar cueing plus automatic classification takes the screen-watching load off a person, which is a real fix for a real failure mode. What remains is an agent free to respond, and classification accuracy stops at stage three. Anduril said in September 2024, announcing its 300th CBP tower, that the systems covered about 30% of the southern border [16].
For a detection rate to support an outcome claim, three records would have to exist and survive: the disposition of each alert, the as-configured boundary at the time it fired, and retention long enough to answer a question raised two years later. All three are record-keeping requirements. Runtimewire's summary of the investigation stated the procurement version of this: contracts "should track response outcomes alongside detections and coverage" [17].
What to watch
- Whether CBP extends tower data retention beyond 30 days or starts logging the disposition of each alert.
- Whether Anduril or CBP release the as-configured virtual boundaries for the three Sunland Park towers.
- Whether the next CBP tower award writes response-time and alert-disposition metrics into the contract terms.