Flock Made Stalking a Feature for Nine Years. The Audit Log Was Always Optional.

Flock Made Stalking a Feature for Nine Years. The Audit Log Was Always Optional.

In August 2026, Garrett Langley, CEO of Flock Safety, told a reporter that the company “knew this was a risk nine years ago.” He was talking about the documented pattern of police officers using Flock’s license plate reader network to stalk ex-partners. As of that statement, 28 confirmed cases had been identified across at least 16 states. Officers tracked ex-girlfriends. A sheriff searched for his wife’s vehicle hundreds of times. A police chief used the system 228 times to monitor his ex and her new boyfriend. In one case, an officer held a gun in a woman’s mouth and did a Nazi salute — the Flock stalking was a footnote in a longer list of misconduct.

Here is what Langley’s statement actually means: Flock built a system capable of generating a comprehensive movement history for any vehicle in America, sold access to thousands of law enforcement agencies, and knew from the start that individual officers would abuse it to track people they knew personally. The company’s response was to make the audit log opt-in. For nine years. Only about one-third of departments turned it on.

I want to be precise about what that decision represents, because the press coverage has been framing this as a surveillance controversy with competing arguments about safety trade-offs. It is not that. It is an engineering and product accountability story with a very clear set of facts. When you build a system with known misuse vectors and make the detection mechanism optional, you are making a product choice. You are deciding that the friction added to sales is more important than the harm the misuse causes. That is a business judgment, not an oversight, and it should be analyzed as such.

The New Yorker piece that surfaced on Hacker News today — 387 points within a few hours of posting — frames this as a story about surveillance capitalism and the post-9/11 normalization of mass monitoring. That framing is correct as far as it goes. But it loses the engineering specificity that makes Flock’s situation distinct from generic surveillance criticism. This is not a story about whether surveillance is bad. This is a story about a company that built a specific system with specific known failure modes, made specific decisions about those failure modes, and is now facing the documented consequences of those decisions while presenting each consequence as a surprise.

What Flock Actually Built

Flock Safety was founded in 2017 by three Georgia Tech alumni — Garrett Langley, Paige Todd, and Matt Feury — around a kitchen table. The origin story Flock tells is that a DeKalb County detective called Langley to say his homemade camera had helped solve a home break-in, and Langley immediately called his co-founders to tell them to quit their jobs. By July 2026, the company operates in over 6,000 communities across 49 U.S. states and performs over 20 billion vehicle scans per month.

That number deserves emphasis: 20 billion scans per month. The U.S. has roughly 290 million registered vehicles. Flock is scanning the equivalent of the entire national vehicle fleet roughly 68 times every month. The company has raised $950 million in venture capital and carries a $7.5 billion valuation. It employs 1,500 people. Annual revenue is approximately $300 million. It is not a startup experimenting with a product. It is an established company with deep municipal contracts operating at national scale.

The camera product line is more sophisticated than its “license plate reader” framing suggests. The base product, the Falcon and Sparrow, reads plates and uses cellular data to hit the National Crime Information Center database in real time, alerting nearby officers to matches. But Flock also performs what it calls “vehicle fingerprinting” — identifying make, model, color, bumper stickers, dents, mismatched panels, and temporary plates as distinct from permanent ones. This allows investigators to search for vehicles by physical characteristics without a plate number, which is useful for crimes where the plate wasn’t captured but witnesses described the car.

The Condor PTZ cameras automatically track individuals, zooming in on faces, phone screens, and other objects nearby. Flock’s marketing says the cameras “focus on vehicles, not people.” Public records requests obtained by 404 Media showed police were in fact searching by race, gender, clothing, height, tattoos, and perceived political affiliation using the system’s natural language FreeForm search feature. The gap between the marketing claim and the documented use case is not ambiguous.

In August 2026, Wired obtained leaked code and prompts from Flock’s OS Investigate tool, previously known internally as Nightshift. The tool correlates ALPR data with police dispatch logs, arrest records, and commercial databases including Social Security numbers, email addresses, phone numbers, and associated individuals. The leaked prompts describe behavior-pattern searches: “someone visiting three or more retail locations within 3 days” or “someone driving their car within a neighborhood in a given time frame.” The tool includes a “workup” function that Flock describes as a “one-command background check” — name, date of birth, owned vehicles, police records, contact details, relatives, and online accounts.

Flock’s CEO described investigators who use OS Investigate as getting “addicted.” This is not metaphorical praise. It is an accurate description of what happens when you give an investigator a tool that can generate comprehensive profiles on anyone without a warrant and without case-number justification. The tool is useful. That is the problem.

The product line also includes the Raven gunshot detection microphone, which now listens for “human distress” beyond gunshots — screaming, per the product marketing. Flock acquired drone manufacturer Aerodome in October 2024. Their Alpha drone reads license plates from 2,000 feet. The drones can be deployed automatically to GPS coordinates generated by gunshot detection or ALPR alerts. The EFF noted that drones can capture footage from areas “typically inaccessible to a casual patrol officer” — people’s backyards, roofs, and through windows.

The Ring partnership, announced October 2025, would have connected residential doorbell cameras into the Flock law enforcement data feed. Amazon ran a Super Bowl LX ad showing Ring’s “Search Party” function finding missing pets using AI surveillance. The backlash was severe enough that both companies canceled the integration in February 2026. The point is not that the partnership failed. The point is that it was attempted at all — that Flock’s product ambition is to integrate every available camera feed, public and private, into a unified queryable network.

The Audit Log Decision

Here is the engineering decision at the center of this story.

When you build a system that logs all queries — who searched for what plate, when, how many times — you can detect misuse algorithmically. Repeated queries on the same plate by the same officer, outside working hours, with no associated case number, is a detectable pattern. This is not a hard problem. It is a straightforward anomaly detection task on a dataset you already have. The data exists. The computation is simple. The only question is whether the output triggers any action.

Flock built exactly this tool. They called it “Audit Assistance.” It flags repeated queries on the same vehicle and queries made outside working hours. It also requires a criminal case number for each query — which means it creates a mandatory justification trail for every search. The product works. The question was never whether Flock could detect misuse. The question was whether they required departments to turn the detection on.

They did not. For nine years, audit logging was opt-in. The CEO says they knew this was a risk from the beginning. Approximately two-thirds of police department customers never enabled it.

In August 2026 — after the Washington Post published a lengthy investigation into a case where an officer stalked his ex-girlfriend using the system, and after the Institute for Justice identified 28 confirmed stalking cases — Flock announced that Audit Assistance would be mandatory by the end of 2026. The announcement coincided with the CEO’s acknowledgment that he had known about this risk for nine years.

I have built production monitoring systems. I know what it means when audit logging is optional. It means the default behavior is no record of who accessed what. It means that when a bad actor abuses access, there is no trail unless someone specifically went looking. It means that the 28 confirmed cases are almost certainly an undercount, because the cases that were discovered were found by journalists and private citizens noticing patterns in public records — records that exist only when departments happened to have audit features enabled or when officers were caught through other means.

More specifically: opt-in audit logging in a law enforcement context is a particularly strange choice because the people administering the system have strong incentives to not look closely. Internal affairs investigations are resource-intensive and politically uncomfortable. A department that has not enabled audit logging has no audit data to review, which means it cannot find misuse it is not looking for. The opt-in design creates a plausible-deniability structure for every department that chose the default. They did not know, because the system was configured to not show them.

There is a direct analogy in software security. When security teams talk about “security by obscurity,” they mean relying on attackers not knowing where to look rather than building actual defenses. Flock’s opt-in audit logging is the internal equivalent: relying on abusers not being noticed rather than building actual detection. Every security engineer knows this fails eventually. Flock’s CEO knew it too, by his own statement. He made it opt-in anyway.

The Scale of What Was Documented

The Institute for Justice’s August 2026 review covered media reports since 2024. That is a two-year window. The 28 confirmed cases represent incidents that were reported, investigated, and made public. Consider the structure of the underreporting problem:

A Medford, Oregon, officer searched Eugene’s Flock data. The documented reason, from internal logs: “hehehe.” This case exists in the record precisely because Eugene happened to have audit logging enabled and a journalist filed a public records request. Without those two coincidences, it does not appear in any count.

Let me list some of the confirmed cases so the scale is concrete rather than abstract:

A Sedgwick, Kansas, police chief used Flock license plate readers to track his ex-girlfriend and her new boyfriend’s vehicles 228 times over more than four months. A Kechi, Kansas, police lieutenant used the system to track his estranged wife. A Bonner Springs, Kansas, detective used it to stalk his wife. A Sumter County, Florida, detective falsified records to search for her husband’s ex-wife. The Braselton, Georgia, police chief was arrested by the Georgia Bureau of Investigation for using Flock to stalk and harass several individuals. A Pasadena, Texas, sergeant used Flock to “track and stalk a female officer.” A supervisor with the Conyers, Georgia, police department used it to track her domestic partner. Two Greer, South Carolina, officers were terminated for misuse. A Niceville, Florida, officer used it to track a fellow officer and that officer’s spouse. A Milwaukee officer tracked a woman he was dating and her ex-partner nearly 180 times over two months. A Louisville, Kentucky, officer was charged with multiple felonies after allegedly using the city’s system to track an ex-partner and her friends hundreds of times over two months. Five Albany, Georgia, police officers were arrested for using Flock for purposes unrelated to law enforcement.

The Brevard County, Florida, case is worth specific attention because of how it was characterized. The officer involved was named “Deputy of the Year” by his department. 404 Media reported that he used Flock to stalk his ex-girlfriend, held a gun in her mouth, did a Nazi salute, and engaged in documented racism and abuse over an extended period. He resigned before being terminated when an internal affairs investigation was opened. The Flock misuse was one item in a misconduct pattern that extended years. He was “Deputy of the Year” while this was happening.

In each confirmed case, the number of tracked queries — 228 times, 180 times, “hundreds of times” — reflects what was recoverable from whatever logs existed. Where no logs existed, the count is zero in the official record regardless of how many actual queries occurred.

The False Arrest Problem Is Separate and Also Structural

The stalking cases share the news cycle with a different category of documented harm: Flock’s plate recognition errors have led to wrongful arrests, detentions at gunpoint, and at least one woman being jailed for 13 days in solitary confinement.

In Roseville, California, a local police department analysis found that 71% of stolen-vehicle alerts Flock sent in 2023 and 2024 were incorrectly identified plates. One thousand and seventeen of 1,427 alerts were wrong. Flock’s response was that Roseville had a “particularly unique deployment” with older hardware placed higher than recommended. The company refused to comment on whether it independently tests camera accuracy.

The Volusia County, Florida, case illustrates what a false positive looks like at the human level. A 23-year-old woman was arrested and held for 13 days, including time in solitary confinement, for a hit-and-run that killed three people. A Flock camera misidentified her vehicle’s color and misread her plate, flagging her as a suspect. The arrest report cited damage on her vehicle “consistent with the accident.” Photo evidence showed her vehicle without damage. A different driver, in a different-colored car of the same model, was later arrested and charged with the actual crime. The 23-year-old spent 13 days in jail, including time in solitary, for a collision she was not involved in, because an automated system made two errors — color and plate — and no human review step existed between the camera alert and the arrest.

A San Diego man was jailed for nearly a month after Flock mistook his plate for one at the scene of an attempted carjacking. Two Morristown, Tennessee, grandparents were detained at gunpoint while their three-year-old granddaughter watched from the car — a Flock camera read an “O” as a “0”. Toledo, Ohio, police detained a driver at gunpoint, deployed their dog on him, and jailed him for hours after a camera read a “7” as a “2”.

A 2021 study by surveillance research firm IPVM found a 10% error rate in Flock’s camera output. Flock responded by halting sales to IPVM and disputing the accuracy of their findings. The company has not published its own accuracy figures. It has not allowed independent third-party accuracy testing. When Business Insider reported the Roseville 71% error rate in 2026, Flock cited deployment conditions rather than addressing the underlying accuracy question.

This matters for a specific engineering reason: the system is designed to match license plates against watchlists and immediately alert nearby officers. There is no human review step between the camera scan and the alert. A 10% error rate on a system doing 20 billion scans per month means 2 billion erroneous reads per month. Almost all of these produce no visible consequence — the erroneous plate is not on any watchlist, no alert fires, the person drives on. But some fraction match watchlist entries. Those generate traffic stops. Traffic stops involving stolen-vehicle alerts, in current law enforcement practice, frequently involve officers approaching with weapons drawn.

The “works on my machine” defense Flock deployed for Roseville is a red flag in any engineering context. When a system produces bad outputs in production and the vendor’s explanation is that the specific deployment was unusual, the correct engineering response is to define what a standard deployment looks like and publish accuracy metrics for it. Flock has not done this. They have disputed specific findings while declining to provide the data that would allow independent assessment.

The Infrastructure Capture Problem

There is a dynamic in Flock’s business model that I have not seen adequately analyzed in the coverage. Flock owns and operates all of its devices and leases them to customers. This is a recurring revenue model that the company frames as a service offering — customers don’t manage hardware, they pay a subscription and get surveillance infrastructure. The problem with this model is that it creates a structural conflict of interest when a customer wants to exit.

When a city ends a contract, the cameras do not automatically come down. Flock has a “standard offboarding process” and “mid-term contract cancellations follow a separate process.” In practice, this means cities that vote to end Flock relationships find themselves negotiating with the company about when the infrastructure will actually be removed.

In Verona, Wisconsin, the city voted in November 2025 not to renew its contract. The cameras remained operational for months after the contract expired despite repeated requests for removal. The city eventually covered the cameras with trash bags. Flock objected to the city removing the cameras themselves — the company’s position was that the cameras are Flock property and only Flock can remove them. The cameras were eventually removed in February 2026, after the city had been publicly bagging them. Flock stated the cameras could be “remotely deactivated at any time” — a claim that raises the question of why they were not deactivated when the city first requested removal.

In Cleveland, Ohio, the cameras remained on and active after a 3-1 city council vote against renewal. Police continued using them. A police sergeant told local media the cameras would remain operational “at minimum” until a council meeting two weeks later allowed for “secondary review.” This is worth being clear about: a democratically elected city council voted 3-1 to end the surveillance contract, and the police department continued operating the surveillance system anyway while seeking a second vote that might reverse the outcome.

In Roanoke, Virginia, 30 of 41 Flock devices installed had been placed in unapproved locations — the city council had approved 75 installations, but 30 of the 41 actually installed were at addresses not on the approved list. In Littleton, Massachusetts, the town alleges Flock reactivated cameras without notifying them after the town had temporarily disabled them.

The operational pattern here is that Flock has designed its business model so that exiting requires negotiation with the company rather than simple contract termination. You cannot turn off the cameras yourself. You cannot remove them yourself. You must go through Flock’s offboarding process on Flock’s timeline. This is a deliberate lock-in mechanism in an industry — law enforcement technology — where lock-in is particularly consequential because the vendor has physical access to public infrastructure and legal relationships with police departments that city councils do not fully control.

The Pattern of Misleading Claims

The false statements documented against Flock extend beyond disputed accuracy figures. Oshkosh, Wisconsin, rescinded its contract after finding that Flock representatives had made “false statements” to the city council. The ACLU issued an explicit denial after Flock falsely claimed to have partnered with them in designing their ALPR system — a claim a Flock representative made before a city council meeting in Urbana, Illinois. The ACLU’s statement: “Neither the ACLU nor any of our affiliates have ever partnered with Flock Safety or worked with them to design any ALPR system.”

When Johnson County, Texas, used Flock cameras to investigate a woman who had a self-managed abortion — querying 83,000 cameras nationwide including cameras in states where using such data in abortion investigations is explicitly prohibited — Flock characterized the investigation on their website as a “missing person investigation.” The sworn affidavit from the detective described an abortion investigation. The company’s characterization directly contradicted the sworn legal document.

In August 2026, after 404 Media published detailed reporting on this case, Flock’s CEO Garrett Langley called Ohio law enforcement officials and told them 404 Media’s reporting was “entirely false.” He argued the woman could not have been under investigation for an abortion because abortion is legal in Texas. This argument ignores that local authorities explicitly discussed whether they could charge her with a crime, and that the investigation included evidence gathering and interviews about the abortion regardless of the ultimate charging decision. 404 Media documented the specific falsehoods in Langley’s characterization in a follow-up piece.

Langley also previously called Will Freeman, the activist who created DeFlock — a crowdsourced searchable map of Flock camera locations — “terrorists” and compared the project to Antifa. He later apologized and called his comments “a mistake.” The DeFlock project received a cease and desist letter from Flock in January 2025. The project continued and now lists dozens of jurisdictions that have ended contracts or deactivated cameras.

A senior Flock publicist falsely claimed on LinkedIn that the company had partnered with the ACLU of New Mexico. The ACLU denied this as well. These are not mischaracterizations or spin. They are factually false statements, documented with primary sources, made by company representatives to city councils, law enforcement officials, and the press.

The Norfolk Ruling and the Fourth Amendment Question

In June 2024, a Norfolk, Virginia, Circuit Court judge ruled that collecting location data from Flock ALPRs constitutes a search under the Fourth Amendment and cannot be used as evidence when collected without a warrant. The ruling applied existing doctrine from United States v. Jones (2012), which found that long-term location tracking via GPS constitutes a Fourth Amendment search, and extended it to ALPR databases that create comparable movement records. The judge explicitly likened ALPR location databases to tracking devices.

This ruling has not propagated nationally. Flock continues operating in 49 states. The constitutional question — whether a network performing 20 billion scans per month, creating a searchable record of every vehicle’s movement history, constitutes a search requiring individualized suspicion — has not been resolved at the federal level. No federal circuit court has ruled on it. The Supreme Court has not taken up the question.

I think the Norfolk ruling got the analysis right. The Jones doctrine rests on the aggregate surveillance problem: short-term location monitoring may not constitute a search, but long-term comprehensive tracking of a person’s movements reveals a “detailed picture of their associations and activities” that the Fourth Amendment was designed to protect. A database of 20 billion scans per month is not short-term monitoring. It is a comprehensive movement record for every vehicle in every covered jurisdiction, available for retroactive querying without a warrant.

The reason this has not generalized is that Fourth Amendment doctrine moves incrementally through specific cases, and courts have been slow to apply existing doctrine to network-scale surveillance. Each individual Flock camera captures a license plate at a specific moment. The constitutional question is about the aggregate — the database, not the camera. That aggregate question requires a court willing to look past the individual tool to the system it creates. The Norfolk court did. Most have not, yet.

In California, state law explicitly prohibits sharing license plate reader data with federal agencies. In 2025, multiple California law enforcement agencies were reported to have shared the data with federal authorities anyway. Johnson County, Texas, queried 83,000 Flock cameras nationwide — including cameras in states where collecting such data for abortion investigations is explicitly prohibited by law — without any apparent mechanism to restrict that cross-jurisdictional query. The data sharing architecture appears to have no enforcement mechanism for legal restrictions on use.

What the New Yorker Piece Gets Right and Misses

The New Yorker framing — “a closely surveilled world with no exit” — correctly identifies the ambient result of the Flock network. If you drive anywhere in a city with Flock cameras, your presence is logged. There is no consent mechanism. There is no opt-out. The framing also correctly situates Flock within the post-9/11 normalization of mass surveillance infrastructure, where safety arguments have consistently been used to shift the burden of proof onto privacy rather than onto surveillance. The company’s language — framing privacy as a “trade-off” for safety — is exactly the rhetorical move the post-9/11 security apparatus normalized two decades ago.

What that framing underweights is the engineering specificity. This is not primarily a values debate about privacy versus safety. It is a specific set of engineering decisions that were made, over time, at a company with $950 million in funding and a $7.5 billion valuation, decisions that had predictable and documented consequences. The audit log was opt-in. The cross-jurisdiction data sharing had no enforcement controls for legal restrictions. The accuracy claims were not subject to independent verification. The camera removal process was designed to disadvantage cities that wanted to exit. The OS Investigate tool aggregates commercial personal data without a warrant requirement. None of these are inevitable features of a license plate reader system. They are choices, and the choices systematically favored growth and revenue over documented harms.

The LAPD ended its Flock contract in July 2026, citing “serious concerns” over civil liberties and privacy. Denver ended its contract and switched to Axon. Austin organizers successfully canceled the city’s contract after a campaign by the EFF. DeFlock now lists dozens of jurisdictions that have ended or deactivated Flock systems. The company’s response to the vandalism of cameras — which has become common enough to be a documented business risk — was to hide the cameras inside the backs of speed limit signs. They are engineering against the public’s ability to locate and remove the surveillance infrastructure.

What Comes Next

Based on the trajectory, several outcomes are predictable with reasonable confidence.

Mandatory audit logging will reduce but not eliminate the stalking cases. Audit logging detects patterns only if someone looks at the flagged outputs. The 13 states that have already made auditing mandatory will see better detection rates than the 37 that have not. The 8 states that have made misuse a standalone crime give prosecutors a cleaner charge than trying to fit Flock stalking into existing harassment or computer fraud statutes. The nationwide pattern will be uneven for years, with documented cases concentrated in states with mandatory requirements and legal consequences.

The federal constitutional question will reach a circuit court within three years. The Ninth Circuit, covering California, and the Fourth Circuit, covering Virginia where the Norfolk ruling originated, are the most likely venues. Whether the resulting ruling applies the Jones aggregate doctrine to network ALPR systems will depend heavily on the specific panel composition. A circuit split is more likely than a clean national resolution. That split gives the Supreme Court grounds to grant certiorari, probably before 2030.

Flock will not go out of business. The network effect of 6,000 municipalities is a genuine moat. Switching costs for law enforcement technology are substantial — new procurement process, new training, loss of historical database access, renegotiation with every connected agency. What I expect instead is a consent decree or settlement with a state attorney general — New York, California, and Illinois are the most likely — that includes mandatory accuracy disclosure requirements and cross-jurisdiction data sharing restrictions with actual enforcement mechanisms. This will be announced by Flock as a transparency and accountability initiative.

The OS Investigate tool is the litigation target that will actually move doctrine. Searching for “anyone who visited three or more retail locations within three days” is not a search for a criminal suspect. It is a population-level dragnet that produces a list of people to investigate. A court applying the Jones aggregate doctrine consistently will find that this requires either a warrant or a showing of individualized suspicion. A court that prioritizes law enforcement utility over Fourth Amendment text will find a way around it. We will find out which kind of court we have when the first challenge reaches a circuit panel.

The cameras are in 49 states. They scan 20 billion vehicles per month. The audit log is still not mandatory until the end of 2026. The CEO knew the risk nine years ago. The product that aggregates ALPR data with commercial personal records and Social Security numbers is called OS Investigate, it is already in active use with law enforcement partners, and the full prompts and capabilities were leaked through a publicly accessible login page. The cameras that Flock installs without consent, operates past contract expiration, and hides inside road signs to avoid vandalism are scanning your vehicle today in 49 states, generating a movement record, and making it available for retroactive query by any officer in any connected jurisdiction without a warrant. That is the product. That is what $950 million in venture capital built. The audit log was opt-in because making it mandatory would have complicated the sales process. The CEO said so.