FlockFacts

Flock Safety: A Documents-First Dossier

Purpose. Every claim below is tied to a specific document at a specific URL. Sources are separated by what kind of record they are — a state auditor's finding and an advocacy group's white paper are both useful, but they are not the same class of evidence and are never mixed here. Items that failed verification are quarantined in §4 and must not be cited.

Standing caveat on scale figures. Flock's camera count is reported inconsistently across confirmed sources: ~90,000 cameras in 5,000+ jurisdictions per the Tan v. San Jose complaint (ij.org), ~120,000 per Cybernews' summary of the WIRED reconstruction (cybernews.com) and the ACLU (aclu.org). Use a range, not a point estimate.

1. The strongest documentary spine

Ordered by how hard each is to rebut.

New — Flock's own patents (added 2026-08-22, verified on Google Patents). These sit at the top of the spine because a granted patent is the one document class where a company must truthfully describe what its technology does, under oath to the USPTO, to obtain a legal monopoly on it. They directly contradict Flock's public line that its technology "cannot recognize, identify, or track individuals." Full exhibit list with verbatim quotes: documents/flock-patent-exhibits.md.

0a. Flock patents biometric classification of people — sex, race, height, weight — plus "face recognition data points" and cross-camera person tracking. US Patent 10,795,933 B1 (granted 2020-10-06, Flock Group Inc), the parent of US11416545B1: patents.google.com. PDF of the child at documents/US11416545B1-flock-object-query-patent.pdf. Hard to rebut: verbatim from the shared specification — "classify the identified object ... the object is a human, is male, is wearing a jacket, and is about 5′11″ and 190 pounds." The spec also recites matching "face recognition data points" and reconstructing a person's and a vehicle's path across many independent cameras. The child patent (US11416545B1, 2022-08-16) adds the explicit "male, female, race" classifier. This is Flock's own sworn description of the product.

0b. Flock patents an on-device detector whose disclosed "identifiable object" expressly includes a human. US Patent 10,559,200 B1 (2020-02-11): patents.google.com. Spec: an identifiable object is "any object ... whether it be a human, an animal, [or an] automobile." Moderate weight — granted claims center on vehicles/plates; the human-identification scope is in the specification.

0c. The "person" object class is still active in Flock's newest cameras (2025). US Patent 12,322,186 B1 (2025-06-03): patents.google.com. The detector is "trained to detect and localize license plates ... as well as other objects, such as vehicles ... and/or people." Proves the capability is current, not a legacy artifact. Moderate weight — claims are illumination-hardware focused; person detection is in the spec/figures.

1. A state's own audit found Flock let U.S. Customs and Border Protection into Illinois cameras in violation of state law — and the state ordered it shut off. Illinois Secretary of State Alexi Giannoulias, audit finding and news release: ilsos.gov · PDF original: ilsos.gov/content/dam. Hard to rebut: this is a constitutional officer's own audit of 12 sampled agencies, published by the state, with an enforcement order attached. The same release documents that Flock ran a federal pilot its own leadership was unaware of, lacked data-sharing safeguards, and paused federal pilots nationwide as a result — an admission of internal control failure, not an outside allegation.

2. A city government terminated its Flock contract and went dark, citing Flock's own admission. City of Evanston, Illinois: content.govdelivery.com. Hard to rebut: a municipality's official release, on the record, calling the audit findings plus "Flock's admission that it failed to establish permissions ensuring local compliance" "deeply troubling." All 19 cameras deactivated; termination effective 2025-09-26. A customer's own words carry weight no critic's can.

3. A state auditor found 99.9% of stored ALPR images were of people not suspected of anything — and no audited agency had a legally compliant policy. California State Auditor, Report 2019-118: auditor.ca.gov. Hard to rebut: independent statutory auditor, published methodology. Of 320 million ALPR images LAPD stored, 99.9% were of vehicles on no hot list; none of four audited agencies had a policy meeting all legally mandated requirements (LAPD had none at all); three shared images with hundreds-to-thousands of entities without verifying they were public agencies with a right or need to access; and agencies had conducted little to no auditing of user searches. Note: this is an ALPR-sector finding, not Flock-specific — cite it for the technology and the oversight vacuum, not for Flock's product.

4. A police department's own analysis found 71% of Flock's stolen/felony alerts to it involved a misread. Business Insider, from hundreds of pages of Roseville PD records: businessinsider.com · department confirmation on the record by Lt. Nick Gaines: sacbee.com · Flock's own root-cause deck delivered to Roseville, obtained by records request: documentcloud.org. Hard to rebut: three independent legs — the department's internal analysis, its spokesman confirming the number publicly (1,011 of 1,427 alerts), and the vendor's own presentation acknowledging the failure modes (blurry distant plates, occlusion by trees and frames, N/V and 1/4 character confusion). ⚠️ Contested: Roseville PD asserts none of those 1,011 misreads produced a contact or arrest (sacbee.com). Concede that point and argue the base rate, not the outcome.

5. A federal court held that a city-wide warrantless Flock dragnet states a valid Fourth Amendment claim. Order denying motion to dismiss, Schmidt v. City of Norfolk, E.D. Va., Chief Judge Mark S. Davis: ij.org. Docket history independently verifiable at the University of Michigan Law School's Civil Rights Litigation Clearinghouse: clearinghouse.net (a law-school research repository, not a government record). Hard to rebut: an Article III judge's signed order. Companion order shows Flock itself moved to intervene to defend its product's constitutionality and was denied — the company's own filing establishes its direct legal stake: ij.org. ⚠️ See §4: the later summary-judgment ruling in this case went for the City, and the appeal is pending. Do not describe Norfolk as a win.

6. Sworn federal complaint: 474 cameras in one city, searched ~2.5 million times in six months, no warrants. Tan v. City of San Jose, class action complaint, N.D. Cal.: ij.org. Hard to rebut: ~15,000 searches per day with no warrant, probable cause, or supervisor approval, pleaded in federal court. A parallel state-constitutional case is live from EFF and ACLU of Northern California: eff.org. Allegations in a complaint are allegations — say so.

7. Flock's own login pages served 450+ unauthenticated files describing an AI tool that searches people, not plates. WIRED: wired.com · independent security-press confirmation: cybernews.com. Hard to rebut: the artifacts came from Flock's own infrastructure, and the reconstruction directly contradicts Flock's public claim that its technology "cannot recognize, identify, or track individuals." See §3.

8. A U.S. Senator put "see no evil" on Flock's record in writing. Sen. Ron Wyden, letter to Flock Safety: wyden.senate.gov. Hard to rebut: official Senate correspondence citing open-records evidence, and referencing Flock's July 25, 2025 agreement to implement additional privacy protections for Oregonians to prevent abuses — i.e., a remedial commitment the company itself made.

9. An academic study of 31 agencies' audit logs found Border Patrol reaching networks that never authorized it. University of Washington Center for Human Rights: jsis.washington.edu. Hard to rebut: records-based, university-published, method stated. At least 8 Washington agencies enabled 1:1 sharing with Border Patrol in 2025; Border Patrol had apparent back-door access to at least 10 more that never authorized it (May–Aug 2025); local agencies ran side-door searches for ICE/CBP in violation of Keep Washington Working. It also records that discrepancies between Flock's audits and other records make independent verification impossible, and that Flock alone among major ALPR vendors refuses independent security testing. And the same center predicted this in 2022, before it happened: jsis.washington.edu.

10. Illinois data was searched for immigration enforcement and for a woman who had an abortion — by out-of-state police, referred to the Attorney General. Capitol News Illinois: capitolnewsillinois.com · Daily Herald: dailyherald.com. Hard to rebut: 46 out-of-state agencies blocked by state order; Mount Prospect's system searched 262 times for immigration reasons in roughly three and a half months; a Texas sheriff searched it regarding a woman who had an abortion; Mount Prospect's own police chief stated on the record that the searches were illegal, and disclosed that Flock's opt-in "National Lookup" let outside agencies search its data with no data-sharing agreement. Independently fact-checked by Snopes against public records: snopes.com.

11. Misuse is a criminal pattern, not anecdotes: at least 20 arrests in one state alone. Reason: reason.com · named cases in 12+ states: newsweek.com. Hard to rebut: named defendants with named dispositions — Georgia DA investigator Andrae Wright (60+ unauthorized queries in three weeks), Habersham County Deputy Christian Brewer, former Costa Mesa officer Robert Jay Josett (guilty plea), former Kechi, KS Lt. Victor Heiar (convicted), Charlotte-Mecklenburg Officer Seth Elliott. Also the accountability gap: Sedgwick, KS Police Chief Lee Nygaard tracked his ex-girlfriend 228 times over four-plus months, admitted it, got 18 months' probation and lost certification — and faced no criminal charges: local12.com.

12. Four distinct security and disclosure failures, documented separately. Cameras exposed to the open internet with no authentication (at least 60 Condor PTZ units; reporters tracked themselves through the feeds): 404media.co · officers' stated search reasons and some searched plates indexed publicly by DuckDuckGo and Bing: 404media.co · government-user credentials found in infostealer malware dumps and Flock accounts for sale on a Russian cybercrime forum: 404media.co · a redaction error in audit logs released under public-records law exposed details of millions of surveillance targets and active investigations, after which Flock threatened the aggregating researcher's hosts and cut back its own audit-log transparency: 404media.co. Hard to rebut: four independent failure modes — device exposure, search-engine indexing, credential hygiene, and product-level redaction — plus a documented retaliation-and-retrenchment response to being caught.

2. By evidence class

Government records

Produced by a government body about its own or a vendor's conduct. Highest evidentiary weight.

Court filings

Signed orders carry judicial weight. Complaints are allegations — label them as such.

Published exposure research

Records-based analysis by journalists, academics, and researchers. Method is stated; the underlying documents are public.

Federal & cross-jurisdictional access

The "side door": local cameras, federal purposes.

Accuracy failures & wrongful harm

Advocacy dossiers

Useful for framing, model policy, and organizing. Never cite as a government finding.

3. The OS Investigate / Nightshift angle

This is the newest thread and the one that changes the nature of the argument: from "a camera network reads plates" to "a search engine for people, built on those cameras."

What was found. Flock's own login portals served unauthenticated visitors a cache of 450+ front-end files describing the authenticated OS Investigate interface (formerly Nightshift). Reporters reconstructed the tool without holding an account — wired.com. Independently summarized in the security press — cybernews.com.

What the reconstruction shows the tool reaches. 69 preloaded police prompts and 45 AI tools spanning plate scans, arrest records, case files, 911 dispatch logs, ballistics, and commercial identity databases carrying Social Security numbers, dates of birth, phone numbers, and relatives — cross-referenced against roughly 120,000 cameras (wired.com, cybernews.com).

Why it matters for the argument, in four steps:

  1. It contradicts Flock's own public claim. Flock has publicly stated its technology "cannot recognize, identify, or track individuals." The reconstructed toolset joins plate scans to arrest records, case files, and commercial identity data — wired.com. The contradiction is between the company's marketing and the company's own served code.
  2. The search key stops being the plate. Secondary coverage of the WIRED work reports that the platform can surface drivers from movement patterns alone — no plate, no name — techtimes.com. ⚠️ This behavioral-search characterization comes from secondary aggregation, not from the WIRED text as confirmed. Attribute it as such, or omit it, if the audience is hostile.
  3. Prompt-based search removes the friction that every guardrail depends on. Every existing safeguard — a "reason" field, a search-term filter, supervisory review — assumes a query names a target. Natural-language prompts over a joined dataset make a search cheap and unbounded; ACLU's Chad Marlow and former officer Noel Pichardo frame this as a "fishing expedition" risk (cybernews.com). That connects directly to the Wichita lieutenant's concession that per-search purpose enforcement is "not even possible" (reason.com) and to the documented defeat of Flock's own reason-field tool by officers typing "hehehe" (aclu.org).
  4. The exposure is itself the security finding. A vendor whose login pages hand 450+ files describing a privileged law-enforcement interface to anonymous visitors is the same vendor that left 60+ PTZ cameras on the open internet (404media.co), indexed officers' search reasons into DuckDuckGo and Bing (404media.co), shipped a redaction failure that unmasked millions of surveillance targets (404media.co), and — per UWCHR — refuses independent security testing of its devices, alone among major ALPR vendors (jsis.washington.edu).

Adjacent, and worth knowing: Flock reversed a plan to incorporate hacked/breached database content into its Nova people-search product after 404 Media's reporting and internal employee pressure — 404media.co. The reversal is evidence the plan existed.

4. Could not verify — do not cite

Each item below was fetched and checked. Some are outright unusable; most are real documents with overstated claims attached. The document is usually fine; the claim about it is not. Where a corrected version is safe to use, it is stated.

Dead link

Court documents — real, but the claims attached were wrong

Oversight correspondence

Journalism — real articles, inflated or inverted claims

5. Gaps — what would strengthen the case but was not found

These are the specific records worth requesting, subpoenaing, or watching for. Each names something that plausibly exists.

Federal oversight documents (none obtained)

Litigation records (docket numbers missing)

Unreleased or partially released audits

Vendor records that would be decisive