CBP’s 2019 pilot breach began with a vendor’s copy made to improve performance
Traveler images stolen from a CBP subcontractor in 2019 were copies it had taken from a border pilot without authorization; the prime contractor’s investigation said it took them to improve performance. Agencies piloting biometric tools should design against a vendor’s appetite for real data.
Looking Back revisits past weeks one at a time, up to the most recent, and asks what it meant for the banks, health systems and public agencies we serve. The series was written in 2026, with the benefit of hindsight, and each article says which week it looks back on.
Two-minute summaryThe dek and the first sentence of each section, quoted from the article
Traveler images stolen from a CBP subcontractor in 2019 were copies it had taken from a border pilot without authorization; the prime contractor’s investigation said it took them to improve performance. Agencies piloting biometric tools should design against a vendor’s appetite for real data.
- The week According to the Department of Homeland Security’s Office of Inspector General, the Washington Post and the New York Times both released articles on June 10, 2019, about a cyber attack involving traveler images from U.S. Customs and Border Protection, and both highlighted that sensitive information had been stolen and placed on the dark web.
- The device: data left at the edge, unencrypted, with an open port The pilot’s data flow, as the inspector general describes it, began at contractor-owned cameras that captured vehicle occupants’ faces and sent them to a lane security controller, which passed image packages to a Vehicle Face System database on CBP’s network for analysis of photo quality and matching accuracy.
- The access: maintenance tickets that opened the data Perceptics staff did not break in.
- The purpose: a vendor that wanted the data to improve its product The inspector general reports that, according to documentation from a Unisys investigation, Perceptics downloaded the images to improve performance.
- The report: a prime that did not tell the customer The fourth failure came after the theft.
- What the attacker took besides faces According to the inspector general, the June 10 articles highlighted that sensitive information had been stolen and placed on the dark web.
- Public trust as a program risk The inspector general devoted a section of its report to the damage the breach might do to public trust.
- A pilot plan built from the report An agency planning a pilot of a biometric or machine learning system with a vendor can turn the inspector general’s findings into a short set of conditions to meet before the first record is captured.
- What the pilot label changed, and what it did not The Vehicle Face System was a pilot.
- What the response got right It would be unbalanced to describe only the failures.
- The strongest objection: CBP did what was required The inspector general’s report records CBP’s own position.
- How the findings apply by kind of agency Agencies piloting biometric or computer vision systems.
- With hindsight The inspector general’s report, “Review of CBP’s Major Cybersecurity Incident during a 2019 Biometric Pilot” (OIG-20-71), was dated September 21, 2020.
The week
According to the Department of Homeland Security’s Office of Inspector General, the Washington Post and the New York Times both released articles on June 10, 2019, about a cyber attack involving traveler images from U.S. Customs and Border Protection, and both highlighted that sensitive information had been stolen and placed on the dark web. The inspector general’s report also quoted the chairman of the House Committee on Homeland Security on the breach, citing a June 10 article in The Hill about the committee’s plans to hold hearings on DHS’s use of biometric information. The inspector general also recorded that a U.S. senator called on DHS that month to halt its use of facial recognition technology after CBP confirmed the breach had exposed images of travelers and their vehicles.
The attacker had not taken the images from a CBP network. They came from the corporate network of Perceptics, LLC, a subcontractor that had installed its facial image capture equipment for a CBP pilot called the Vehicle Face System at the Anzalduas, Texas port of entry. Unisys Corporation was the prime contractor. The pilot, begun in 2018, was testing whether cameras could capture the faces of volunteer vehicle occupants at speeds under 20 miles per hour in both inbound and outbound lanes, and whether CBP’s Traveler Verification Service could match them against a gallery of recent travelers. According to the inspector general, the images were not used to verify identities or create border crossing records. The pilot was meant to inform the expansion of biometric verification for people entering and leaving the country by vehicle.
By the time of the news reports, the government’s response was already under way. The inspector general’s account says CBP first learned of the breach on May 24, 2019, and reported it to the Department that day, notifying offices including the DHS Chief Information Security Officer, the Office of Inspector General and the Enterprise Security Operations Center. On June 3, 2019, DHS declared the event a “Major Cybersecurity Incident,” based on the potential impact to the Department’s reputation and demonstrable harm to public confidence. CBP notified Congress of the major privacy incident on June 8, 2019, and CBP stood up a DHS breach response team that included the Under Secretary for Management, the Chief Information Officer, the Chief Information Security Officer and the Chief Security Officer, along with representatives from DHS privacy, partnership and engagement, general counsel, public affairs and legislative affairs offices.
Also in June 2019, CBP temporarily suspended Perceptics from participation in future government contracts, subcontracts, grants, loans and other federal assistance programs. The suspension was taken under the Federal Acquisition Regulation’s rules on suspension and debarment, which the inspector general described as a measure taken to protect the government, not to punish.
Perceptics was not new to CBP. The inspector general noted that before the Vehicle Face System pilot it had worked for CBP as a subcontractor providing license plate reader technology at multiple Border Patrol checkpoints, and that it had performed technical work at air, land and sea ports of entry. The pilot itself was part of CBP’s wider effort to extend biometric matching, already in use for air departures, to people leaving the country by land.
The congressional reaction the inspector general recorded in June 2019 centered on DHS’s use of facial recognition and biometrics. That debate matters, but on our reading it misses what the inspector general found when it reconstructed what happened. The images were stolen from a vendor’s network. They were on the vendor’s network because the vendor’s staff had copied them from the pilot equipment, without CBP’s authorization or knowledge. And according to the prime contractor’s investigation, cited by the inspector general, the vendor downloaded the images “to improve performance.”
That last detail is what makes the case a data governance case as much as a security one. A vendor building a biometric or machine learning product may have a commercial interest in real-world data from its deployments. Real images from a real border crossing, at real speeds, in real light, are the kind of data a vendor may want for making its system better. When the vendor has physical access to equipment holding that data, the incentive and the opportunity meet. The CBP pilot showed what happens when nothing in the design stops them meeting.
We read the inspector general’s report as describing four points at which the copy could have been prevented or caught: the device that held the data, the access the vendor was given, the purpose for which the vendor wanted the data, and the chain of reporting once something went wrong. Each point belongs to a different part of an agency: engineering, operations, contracting and incident response. We take each in turn. We then look at what else the attacker took, at public trust as a risk to the program, at a pilot plan built from the report’s findings and at what the response got right, before asking what the label “pilot” did and did not change. Finally we take the strongest objection to our reading, and set out how the findings apply to different kinds of agency.
The device: data left at the edge, unencrypted, with an open port
The pilot’s data flow, as the inspector general describes it, began at contractor-owned cameras that captured vehicle occupants’ faces and sent them to a lane security controller, which passed image packages to a Vehicle Face System database on CBP’s network for analysis of photo quality and matching accuracy. The images Perceptics took came from a computer system at the test site, connected to the cameras and containing images of drivers and passengers collected during the pilot.
The inspector general found that CBP did not adequately safeguard sensitive data on an unencrypted device used during the pilot. An open USB port allowed Perceptics staff to connect an unencrypted hard drive and download unencrypted biometric images. The report states the principle directly: even though the subcontractor provided the equipment, CBP remains responsible for securing its technology, under the DHS 4300A Sensitive Systems Handbook.
Two design choices combined. The data was stored, unencrypted, on equipment at the test site, and the equipment allowed removable media to be attached. A 2017 DHS privacy policy already classified biometric information as sensitive personally identifiable information and prohibited DHS employees from connecting non-government-issued removable media to DHS equipment or networks. The policy was clear. The equipment did not enforce it.
After the breach, the report says, CBP’s Chief Information Security Officer acknowledged the equipment vulnerabilities at the pilot location, and CBP disabled all USB capabilities to prevent further unauthorized access to pilot data. CBP staff told the inspector general that about four months after the breach they had performed the software updates needed to support encryption on equipment similar to that used in the pilot. CBP then ran forensic security assessments of cameras and biometric technology at other locations. As of November 8, 2019, it had completed onsite evaluations at five locations: four major international airports in the biometric air exit program and a testing facility in Sterling, Virginia. The more rigorous examinations, at three of the five locations, found no traveler biometrics stored on the devices. CBP identified potential security vulnerabilities at four airports conducting similar facial recognition pilots and made ten mitigation and three policy recommendations, one key recommendation being USB device restrictions and enhanced encryption.
For agencies running sensor-based programs, the point is that the edge device is part of the information system. A camera controller, a license plate reader, a kiosk or a body-worn camera dock holds sensitive data, if only briefly. It should be inside the system’s security authorization, with encryption at rest, ports disabled unless needed, and logging of any connection. When a vendor supplies and maintains the device, those controls should be written into the specification and verified at acceptance, not assumed.
There is also a question of how much data the edge needs to keep at all. A camera controller needs to hold an image long enough to send it on. Whether it needs to keep a store of images once they have reached the agency’s database is a design decision. The report’s figure of approximately 184,000 traveler images downloaded from the pilot equipment in conjunction with at least three work order tickets, a number the inspector general says it could not independently validate, shows how much a device at a single test site could yield. A device that forwards each image and then deletes it holds little for anyone to copy. Where a pilot needs a local store, for example to evaluate image quality, the store can be kept small, encrypted and purged on a schedule the agency checks. Data minimization, which agencies apply to what they collect from the public, applies with equal force to where they let copies accumulate.
The access: maintenance tickets that opened the data
Perceptics staff did not break in. They came in through the agency’s own maintenance process. The inspector general found that Perceptics gained access to CBP’s data by submitting work order tickets through the CBP IT help desk, on at least three occasions: August 31, 2018, November 2, 2018, and January 31, 2019. The work requested was maintenance: adjusting the ground loop sensitivity, replacing camera lenses and switching cameras to monochrome at CBP’s request. Once CBP and Unisys approved the tickets, Perceptics personnel did the maintenance, and also used the access to download images. None of the tickets authorized Perceptics to access or download images.
This is the operational heart of the case. The approval process asked whether the vendor should be allowed to service the equipment. It did not ask, or could not enforce, what the vendor could do with the data on the equipment while it was there. A maintenance visit to a device holding sensitive data is access to that data unless something prevents it.
Agencies have several ways to separate maintenance from data access. Equipment can be designed so that the data store is encrypted with keys the vendor does not hold, so that a technician can service the camera without being able to read stored images. Maintenance can be escorted, with an agency employee present who knows what the technician is permitted to do. Sessions can be logged and reviewed afterward, so that a download shows up in a record someone reads. Data can be moved off the edge device promptly and deleted there, so that a maintenance visit finds little to take. Work order approvals can state, in terms, that the work does not include access to stored data, and require the technician to sign an acknowledgment.
These measures are ordinary. We think each would have made copying 184,000 images onto a hard drive harder or more visible. We read the report’s finding that CBP’s information security practices during the pilot “were inadequate to prevent the subcontractor’s actions” as a finding that the maintenance path was also a data path, and nobody had closed it.
The purpose: a vendor that wanted the data to improve its product
The inspector general reports that, according to documentation from a Unisys investigation, Perceptics downloaded the images to improve performance. CBP did not know of or authorize the removal of the data or its storage on the subcontractor’s network. The unencrypted USB hard drive was eventually transported to the company’s corporate office in Knoxville, Tennessee, and uploaded the images to a Perceptics server.
The report also describes a second, older set of data. During its investigation, CBP learned that Perceptics had previously obtained more than 105,000 license plate images from earlier pilots. Those images had been obtained through a CBP-authorized process aimed at improving the license plate reader program, during a 2008 to 2010 contract and a 2016 tactical pilot. But they had been stored on Perceptics’ servers for longer than the permitted one year.
Taken together, these two findings describe a vendor that collected operational data from its government deployments to improve its technology, once with authorization and a retention limit that was not kept, and once without authorization at all. In our judgment, that pattern is a central data governance risk for agency programs that use a vendor’s biometric, computer vision or machine learning system. The vendor’s product gets better with real data. The agency’s data is the best real data available. Unless the contract, the technical design and the oversight all say otherwise, the two will tend to meet.
For an agency, the response starts in the contract. It should say whether the vendor may use agency data to train, tune or test its products, and if so, which data, under what protections, where it may be stored, for how long, and how its deletion will be proven. If the answer is that the vendor may not use the data at all, the contract should say that, and the technical design should make it hard. If the answer is that the vendor may use some data under conditions, the agency should treat that use as a separate data flow with its own privacy assessment, security controls and audit. The 105,000 license plate images show that an authorized use with a retention limit is still a risk if no one checks that the limit is kept.
There is a second contract question that the case raises without answering. When a vendor improves its product using an agency’s data, the improvement lives in the vendor’s software, which it may sell to others. An agency should decide, before the pilot, whether it is willing for its data to improve a product that will be sold elsewhere, and on what terms. That is a policy choice as well as a security one, and it is better made in the solicitation than discovered in an investigation.
The report: a prime that did not tell the customer
The fourth failure came after the theft. Perceptics’ corporate network was hit by a ransomware attack at some point before May 13, 2019. The inspector general recounts that Perceptics received a ransom note demanding 20 bitcoin within 72 hours, threatening to upload stolen data to the dark web, and that Perceptics did not pay. The attacker uploaded more than 9,000 unique files. Besides the traveler images and license plate images, the attacker stole contractual documents, program management documents, emails, system configurations, schematics and implementation documentation related to CBP’s license plate reader programs.
The contracts required prompt reporting. The inspector general’s table of violations lists the requirement that all known or suspected sensitive information incidents be reported to the headquarters or component security operations center within one hour of discovery, as the DHS 4300A handbook requires, a requirement written into the Unisys contract with CBP. Yet the report says that Perceptics and Unisys both defied contractual obligations and DHS requirements for immediately reporting privacy incidents. “Unisys chose not to inform CBP immediately of the data breach.” CBP found out from a news article, approximately one week after Perceptics had notified Unisys.
That gap shows how a reporting chain can break at the first link above the vendor. The subcontractor told the prime. The prime did not tell the customer. The customer learned from the press. The stolen material included documents about CBP’s systems, so the gap mattered for security as well as for privacy: schematics and configurations in an attacker’s hands can help plan further attacks.
An agency can do several things about the reporting chain. It can make sure the reporting obligation flows down to every subcontract, as the inspector general found the relevant clauses had flowed from CBP to Unisys and from Unisys to Perceptics. It can require subcontractors with access to sensitive data to report directly to the agency, as well as to the prime, so that a prime cannot sit on the news. It can name, in the contract, the agency office to be told and the means. It can also test the chain before it is needed, with an exercise in which a subcontractor reports a simulated incident and the agency measures how long the news takes to arrive and through whom. And it can make consequences real: the report records that CBP suspended Perceptics in June 2019, and notes that a contractor’s failure to comply with an administrative compliance agreement is cause for debarment.
What the attacker took besides faces
According to the inspector general, the June 10 articles highlighted that sensitive information had been stolen and placed on the dark web. The report’s account of what was exposed is longer. Besides the more than 184,000 traveler facial image files and 105,000 license plate images that CBP determined were stored on the subcontractor’s network at the time of the attack, the hacker stole contractual documents, program management documents, emails, system configurations, schematics and implementation documentation related to CBP’s license plate reader programs. More than 9,000 unique files were uploaded to the dark web after Perceptics did not pay the ransom.
For an agency, that second list may matter as much as the first. System configurations and schematics describe how an agency’s equipment is built and connected. Implementation documents show where the devices are, how they talk to agency networks and what software they run. In an attacker’s hands, they are a guide to the agency’s systems. A vendor that supports an agency’s technology may hold this kind of material, and its security then bears on the agency’s security even when the vendor holds no personal data at all.
Privacy rules, breach notification requirements and public concern focus on data about people. System documentation falls outside those categories, but it can be controlled unclassified information in its own right, and its loss can enable later attacks on the systems it describes. An agency’s inventory of what its vendors hold should include design and configuration material as well as records about individuals, and the vendor’s obligations to protect, report and destroy that material should be as clear as its obligations for personal data.
The point extends past the prime contractor. Contractual and program management documents and emails about CBP’s programs were in Perceptics’ files. A subcontractor several steps from the agency may hold copies of documents that the agency regards as internal. Asking each vendor, at the start of a contract and at its end, what agency documents it holds and where, is a simple step that makes the later question of what was exposed answerable.
Public trust as a program risk
The inspector general devoted a section of its report to the damage the breach might do to public trust. It noted the June 10 news coverage, the calls from members of Congress, and earlier congressional caution about biometric programs, and concluded that the incident might make travelers reluctant to let DHS capture and use their biometrics at ports of entry.
For a program that depends on people agreeing to be photographed, that is a mission risk as well as a public relations problem. The Vehicle Face System pilot used volunteers. If people decline to take part because they doubt their images will be protected, the program’s data gets thinner and its value falls.
Agencies can do some things to protect trust before an incident. They can say publicly and clearly who handles the data a program collects, including vendors and partners, and what those parties may and may not do with it. CBP’s published privacy documents were a step in that direction. They can describe how long data is kept and how its deletion is checked. And when an incident happens at a vendor, they can explain what was taken, what was not, and what has changed, without waiting for the press to set the terms. The inspector general’s account shows CBP learning of the breach from a news article, which left the agency responding rather than explaining.
The case also shows the limits of good privacy paperwork. CBP had produced 56 privacy products for the program. Not one of them stopped a technician with a hard drive. Trust rests on the public’s belief that commitments are enforced, and that belief is tested at exactly the points where the Perceptics case failed.
A pilot plan built from the report
An agency planning a pilot of a biometric or machine learning system with a vendor can turn the inspector general’s findings into a short set of conditions to meet before the first record is captured.
- Every device that captures or stores data is inside the system’s security authorization, encrypts data at rest, has unused ports disabled and logs connections.
- Data moves off edge devices to agency systems promptly and is deleted on the device on a stated schedule.
- Vendor maintenance is scheduled, logged and, where data is present, escorted or performed without the ability to read stored data.
- The contract states whether the vendor may use pilot data to improve its products, and if so which data, where, for how long, and how deletion will be shown.
- Subcontractors with data access report incidents to the agency directly, with the office and means named, and the prime’s contract requires it.
- The agency keeps a list of every party that holds pilot data or system documentation, updated when vendors change.
- The pilot’s privacy documents say who handles the data and what they may do with it, and a named official checks that the commitments are met.
These conditions add cost to a pilot. The inspector general found that additional IT security controls in place during the pilot could have prevented the download; the other conditions are our own reading of where the case shows gaps. A program office that cannot meet them for a given pilot should record why, and who accepted the risk. That record is what an inspector general, a congressional committee or the public may ask for if something goes wrong, and it is far easier to write before the pilot than to reconstruct after an incident.
What the pilot label changed, and what it did not
The Vehicle Face System was a pilot. It was testing image capture at speed, at one port, with volunteers. Its images were not used to verify identities or create crossing records. It would be natural for an agency to treat such a test as lower risk than a production system.
The inspector general’s report shows that the label did not change the sensitivity of the data. DHS considers biometric information such as facial images to be sensitive personally identifiable information, and the report applied that classification to the pilot’s images. The images of real people crossing a real border carried the same risk in a test as in operation. In our assessment a pilot may, if anything, carry more risk, because the equipment is newer, the vendor is more closely involved, maintenance is more frequent and the controls of a mature system may not yet be in place.
The report also records what CBP did right before the breach. The Biometric Entry-Exit Program Office and the CBP Privacy Office had produced 56 privacy products during the program’s development, evaluating privacy aspects of the program and explaining how privacy concerns were mitigated, with some published on DHS’s website. Perceptics employees had completed all required training courses, including IT security awareness and rules of behavior training, privacy training, integrity training and privileged user training. All relevant DHS special clauses and Homeland Security Acquisition Regulation clauses had flowed through from CBP to Unisys and from Unisys to Perceptics.
Paperwork, training and clauses were all in place. The data still left on a hard drive. What the pilot lacked was enforcement at the point where the vendor touched the data: a device that would not accept the drive, a maintenance process that separated servicing from data, a contract that addressed the vendor’s interest in the data, and a reporting chain that reached CBP before the press did. For a program office planning a pilot, the useful question is which of its privacy commitments are enforced by the system and which depend on everyone choosing to comply.
What the response got right
It would be unbalanced to describe only the failures. Once CBP knew, the inspector general describes it as taking prompt action to notify the Department and mitigate risks. It reported the incident to the Department the day it learned of it, informing the Chief Information Security Officer, the Inspector General and the Enterprise Security Operations Center. DHS declared a major incident, and CBP notified Congress of the major privacy incident. A breach response team was formed with the Under Secretary for Management, the Chief Information Officer, the Chief Information Security Officer and the Chief Security Officer, along with privacy, legal, public affairs, legislative affairs and partnership staff.
The response team’s actions, between May 24 and October 8, 2019, went to the source. It removed from service all equipment involved in the breach, canceled Perceptics’ employees’ access to CBP systems and data, and required Unisys to terminate its contract with Perceptics. CBP opened its own investigation of Perceptics, which found the older license plate images. It suspended the company. Its privacy office assessed the likelihood of harm to individuals using OMB’s breach guidance, and the Department’s acting chief privacy officer provided that assessment to Congress.
CBP also looked beyond the pilot. Its Chief Information Security Officer began forensic assessments of cameras and biometric technology at other sites to make sure data was not being stored on other endpoint devices. CBP sent a memo requiring all IT contractors to sign statements guaranteeing compliance with contract terms on IT and data security, with supporting documents and answers to a questionnaire titled “Baseline Security Requirements for Securing Sensitive Data.” As of October 11, 2019, it was collecting the attestations.
These steps show an agency that, once alerted, treated a vendor incident as its own. What the case leaves for others is about the period before the alert. The response machinery worked from May 24. The prevention machinery, and the reporting chain that should have triggered the response sooner, did not.
The strongest objection: CBP did what was required
The inspector general’s report records CBP’s own position. In its response to the draft report, CBP maintained that it did what was required to protect the data associated with the pilot, and outlined standard protection measures and contractor requirements meant to protect data collected by the Department. On this view, the breach was caused by a subcontractor’s staff deliberately violating rules they had been trained on and agreed to, and by a prime contractor that failed to report. No set of policies prevents a determined insider.
There is real force in this. The violations were by contractor staff, and the inspector general found that Perceptics staff directly violated DHS security and privacy protocols. CBP had flowed down the clauses, required the training, done privacy assessments and, once it knew, acted quickly to remove equipment, cancel access, require Unisys to terminate its contract with Perceptics, and suspend the company.
But the inspector general did not accept the objection in full, and neither do we. The report’s response was that, although CBP maintained it had done what was required, “the data was still removed without authorization.” It found that CBP did not implement all available IT security controls, including an acknowledged best practice, and that additional controls could have prevented the download. The distinction matters for any agency. Policies and clauses allocate blame after an incident. Technical controls determine whether the incident happens. An agency that relies on the first without the second has a strong legal position and a weak security one.
A second objection is that the harm was limited. After removing duplicate images, CBP reduced its estimate to 100,000 individual images, of which it found 19 posted on the dark web. The stolen images were not linked to other traveler personal information, and CBP’s privacy office, assessing the likelihood of substantial harm to individuals, determined the information taken was of low risk. That is a fair point about individual harm. The inspector general’s concern was broader: that the breach might damage public trust in the government’s ability to safeguard biometric data and make travelers reluctant to let DHS capture their biometrics. For a program that depends on public participation, that is a harm to the mission even if no individual suffers loss.
A third objection is that the controls we describe would not stop a vendor determined to take data. A technician with authorized access to a device can find ways around port restrictions, and a vendor with a legitimate need for some data can exceed it. That is true, and no control set is complete. But the controls change the act. Copying data from an encrypted device without the key, during a logged and escorted session, in breach of an explicit contract term on data use, is harder, more visible and more plainly wrongful than plugging a drive into an open port during routine maintenance. Controls that make misuse difficult and detectable shift both the vendor’s calculation and the agency’s chance of finding out before an attacker does.
How the findings apply by kind of agency
Agencies piloting biometric or computer vision systems. Any agency testing face, iris, fingerprint, gait or license plate recognition with a vendor faces the Perceptics pattern directly. The vendor’s equipment will hold sensitive data, the vendor will need maintenance access, and the vendor will want real data. The pilot plan should address all three before the first image is captured.
Agencies buying machine learning products. The same incentive exists wherever a vendor’s product improves with operational data, whether the data is images, documents, call recordings or case records. The questions of training rights, retention, location and proof of deletion apply even where no camera is involved. Cloud-hosted products add a twist: the data may never leave the vendor’s environment in the first place, which makes the contract and the audit rights the agency’s main tools.
State and local agencies. License plate readers and camera systems are used by state and local law enforcement and transportation agencies, which may be supplied and maintained by vendors. The 105,000 license plate images held past their retention limit are a reminder that retention limits written into contracts need checking. A state or local agency can ask its vendor to certify, periodically, what agency data it holds and where, and test the answer.
Agencies working through prime contractors and integrators. Where a prime integrates several subcontractors’ products, the agency’s visibility into each subcontractor is limited. The Unisys failure to report shows the cost. Direct reporting obligations on subcontractors, and the agency’s right to know who all the subcontractors with data access are, reduce the dependence on one link.
Agencies early in adopting machine learning. An agency running its first pilot of a vendor’s model may not yet have contract language on training data, a process for approving vendor data use, or staff who can check a vendor’s deletion claims. The Perceptics case suggests building those before the pilot, not after. Standard contract terms on data use, adopted once and reused, cost less than negotiating each pilot afresh, and they give program offices a default that protects the agency when no one thinks to ask.
Agencies sharing data with private partners. CBP’s biometric programs involve airlines, airports and cruise lines capturing images on CBP’s behalf. Each partner is another point where the purpose question arises: a partner may have its own uses for images it captures. Rules that forbid partners to keep images for their own purposes, and give the agency the right to audit, address the same risk the Perceptics case exposed.
With hindsight
The inspector general’s report, “Review of CBP’s Major Cybersecurity Incident during a 2019 Biometric Pilot” (OIG-20-71), was dated September 21, 2020. It made three recommendations, and CBP concurred with all three: implement all mitigation and policy recommendations from CBP’s security threat assessments, including USB device restrictions and enhanced encryption; make sure the additional security controls were implemented on relevant devices at all existing Biometric Entry-Exit pilot locations; and establish a plan to routinely assess third-party equipment supporting biometric data collection for compliance with Department security and privacy standards. The report also recorded that CBP had lifted the Perceptics suspension on September 26, 2019, after the two entered into an administrative compliance agreement, and that at the conclusion of the inspector general’s fieldwork Perceptics was no longer working with CBP as a prime contractor or subcontractor.
DHS later put contractor incident reporting duties into the Homeland Security Acquisition Regulation itself. A final rule amending the regulation to safeguard controlled unclassified information was published on June 21, 2023, and took effect on July 21, 2023. Its safeguarding clause requires contractors to report known or suspected incidents involving personally identifiable information or sensitive personally identifiable information within 1 hour of discovery, and all other incidents within 8 hours of discovery. The rule also requires certification of sanitization of government and government-activity-related files, and procedures and capability to notify and provide credit monitoring to individuals whose information was under the contractor’s control at the time of an incident. Contractors must insert the clause in all subcontracts, and require lower-tier subcontractors to do the same, where subcontractor employees will have access to such information, the information will be collected or maintained by a subcontractor, or a subcontractor’s systems will process it.
The biometric program itself moved from pilots to rule. On October 27, 2025, CBP published a DHS final rule, with a request for comments, effective on December 26, 2025, providing that DHS may require all aliens to be photographed when entering or exiting the United States, removing references to pilot programs, and removing the port limitation on collecting biometrics from departing aliens. The rule describes CBP’s business requirements for public-private partnerships, which prohibit approved partners such as airlines, airport authorities and cruise lines, and participating organizations including vendors, systems integrators and other third parties, from retaining the photos they collect for their own business purposes. Partners must immediately purge the images after sending them to CBP and must allow CBP to audit compliance. The rule says any photos of U.S. citizens will be discarded within 12 hours of verification of the individual’s identity and citizenship. Vehicles, the subject of the 2019 pilot, were still not covered at scale: the rule says CBP is still determining the best way to implement biometric entry-exit in some environments, and that its future work may include pilot programs for travelers entering and exiting in vehicles at land ports.
Those business requirements answer the purpose question the Perceptics case raised, for the partners and vendors they cover: a third party may not keep the images for its own use, and the agency may check. The rule’s preamble also responded to a commenter who urged CBP to carry out the recommendations of a 2020 Government Accountability Office report on CBP’s facial recognition programs (GAO-20-568), saying CBP has a minimum acceptable accuracy rate for the program and regularly tracks it, and that CBP concurred with the GAO recommendations and has addressed each one. Whether the audit rights in the business requirements are exercised, and how, is a matter the rule does not report.
Spotted an error? Tell us and we will check it and correct the article where needed.