Flock Safety built a fake police department to test searches on real cameras

Why it matters

Flock Safety's live-camera tests show why surveillance guardrails cannot depend mainly on warnings and audits. Demo accounts need hard separation from production networks before employees or customers can run sensitive searches.

Garrett Langley (@glangley) built Flock Safety around a direct promise: cameras could give police useful evidence while software, audit logs and local oversight kept the resulting surveillance accountable. Records published by 404 Media on September 17th show where that promise broke down during Flock Safety's own demonstrations.

Flock Safety created a fictitious law-enforcement identity called "Flock City PD" and used accounts bearing that name to search live camera networks in Dunwoody, Georgia, Bryan, Texas, and several unidentified jurisdictions, according to audit logs obtained by Dunwoody resident Jason Hunyar. The searches covered people wearing masks, crowds, groups holding signs, political bumper stickers, religious symbols and other characteristics captured by real cameras. The dates of the searches were not established in the report.

Flock Safety told 404 Media that its employees were testing FreeForm search moderation and demonstrating queries that police should avoid. The searches were not customer investigations, according to Flock Safety. Some prompts were blocked, others generated warnings, and searches for "crowd," "person wearing scrubs" and "large group with signs" were allowed.

The distinction matters, although it does not settle the issue. Flock Safety was testing its own controls rather than pursuing a resident as a suspect. It still used live systems containing images of real people to conduct that test. For a founder who has made auditability and customer control central to Flock Safety's pitch, the episode exposes a governance failure inside the product-development and sales process.

From license plates to open-ended visual search

Langley started Flock Safety after break-ins in his Atlanta neighborhood showed him how little investigators could do with grainy conventional camera footage. A 2009 electrical engineering graduate of Georgia Tech, he contacted fellow alumnus Matt Feury and began assembling lower-cost cameras from readily available components. Langley has said an early installation produced evidence that helped police make an arrest within 60 days.

The original product focused on converting passing vehicles into structured evidence: plates, colors, body types and visible damage. Flock Safety has since expanded into video cameras, gunshot detection, drones and cloud software that connects data across participating organizations. Its FreeForm product lets authorized users describe what they want in ordinary language, including clothing, visible accessories and vehicle details. Flock Safety says people searches use observable characteristics rather than facial recognition.

That shift changes the risk profile. A fixed plate query asks whether a known identifier passed a camera. A FreeForm query asks an image-recognition model to decide which unknown people or vehicles resemble a description. The audit logs show searches extending into political expression, religion and assembly, areas where a technically successful match can create a civil-liberties problem.

Flock Safety's moderation system distinguishes among blocks, warnings and allowed searches. In an April 16th explanation of its testing program, Flock Safety said offensive prompts and searches using personal information should return no results. Queries that may implicate First Amendment rights instead produce a warning and can proceed, with the search referred for audit.

A warning is an accountability mechanism rather than a technical prohibition. Its effectiveness depends on the user stopping, the search being sent for audit and an agency enforcing consequences. The Flock City PD records illustrate how much of the product's safety model rests on that chain of human decisions.

The test environment was somebody else's city

Flock Safety has argued that real-world testing is necessary to train visual models and verify moderation. In April, Flock Safety said Dunwoody had given explicit permission to participate in its testing program. Flock Safety compared the process to cybersecurity testing, where a simulated attack is used to check whether a defense holds.

The analogy weakens when the simulated misuse runs against a production camera network. The people and vehicles captured in Dunwoody did not become synthetic test data because a city authorized Flock Safety's access.

That tension had already surfaced before the Flock City PD records became public. In April, 404 Media reported that Flock Safety employees accessed cameras covering a children's gymnastics room, a playground, a school, a pool and a Jewish community center during demonstrations. Flock Safety confirmed the authorized demo access while rejecting claims that its employees were spying on children. Flock Safety and Dunwoody subsequently said the city's cameras would no longer be used for sales demonstrations.

Flock Safety told 404 Media that the Flock City PD environment has since been isolated and can no longer connect to live law-enforcement camera networks. A more recent training webinar showed the account connected to cameras labeled for demo or training use. That is the separation Flock Safety needed from the start: a controlled environment where sales teams and engineers can test dangerous queries without searching residents going about their day.

Langley's accountability bet gets harder at scale

Langley has spent years presenting privacy controls as part of the product rather than a concession attached to it. In 2022, he wrote that Flock Safety's mission was to eliminate crime within a framework protecting privacy and promoting objectivity. At the time, he said users could identify the presence of a bumper sticker without filtering for its political or religious content. FreeForm has since made far more expressive searches possible, leaving moderation to police a capability that earlier product boundaries largely prevented.

Investors have financed that expansion. Flock Safety announced a $275 million round on March 13th, 2025, at a $7.5 billion valuation, led by Andreessen Horowitz. Greenoaks Capital, Bedrock Capital, Meritech Capital, Matrix Partners, Sands Capital, Founders Fund, Kleiner Perkins, Tiger Global and Y Combinator participated. Flock Safety also said annual recurring revenue had passed $300 million after growing 70% year over year. Those financial and growth figures were reported by Flock Safety and have not been independently substantiated.

That capital supports a broader ambition than selling license-plate cameras. Flock Safety is assembling an operating layer for police departments, combining sensors, searchable video, cross-agency data and real-time response tools. The commercial advantage comes from putting more evidence behind one query. Each new data source also increases the consequences of a permissive account, a poorly designed demo or a warning that users learn to click through.

The political cost is already reaching contracts. RuntimeWire reported in May that Bandera, Texas, canceled its Flock Safety agreement after sustained objections to the camera network. The Dunwoody records give local opponents a concrete governance question to take into similar debates: whether cities understand who can access their systems for testing, development and sales.

Jason Hunyar's comments to 404 Media put the concern more sharply. Flock Safety's ability to moderate sensitive searches also proves that the underlying network can capture and retrieve those attributes, he argued. The policy gate belongs to Flock Safety, a for-profit vendor whose software and product scope continue to change.

Flock Safety's decision to isolate Flock City PD addresses the immediate access problem. Langley's harder task is making that separation durable across a growing organization where engineers need test data, salespeople need convincing demonstrations and police customers want fewer limits on useful searches. Audit logs made this episode discoverable. The records also show that transparency works after access has occurred, while sound product architecture can prevent the access entirely.