# A Flock Camera Made 1.6 Million Images. Its Key Stayed on the Device.

**Summary:** A recovered Flock camera logged 1.6 million images in 21 days and kept an encryption key on the device. I think physical capture belongs in every city's threat model.

- Canonical: https://markhuang.ai/news/flock-camera-1-6-million-images-physical-access
- Language: en
- Author: [Mark Huang](https://markhuang.ai/about)
- Published: 2026-09-16
- Section: News
- Tags: Flock Safety, License Plate Readers, Surveillance Technology, Hardware Security, Data Privacy
- Source: [WIRED](https://www.wired.com/story/hackers-flock-camera-data-shows-how-system-works/)
- License: https://creativecommons.org/licenses/by-nc/4.0/

---

![An opened roadside camera reveals its electronics beside layered images of passing cars](https://cdn.markhuang.ai/news/flock-camera-1-6-million-images-physical-access/hero.webp)

*A roadside camera is a computer left in public. Encryption has to survive the day someone gets the computer.*

A hacker collective removed a Flock Safety camera and copied much of its storage. The resulting [WIRED and 404 Media investigation](https://www.wired.com/story/hackers-flock-camera-data-shows-how-system-works/) found logs covering about 21 days, roughly 50,200 photographed vehicles, and about 1.6 million generated images. The reporters also recovered thousands of short videos after the hackers found an encryption key on the device.

I keep coming back to the location of that key. Flock has treated the need for physical access as an important limit on earlier camera vulnerabilities. I see it as the threat model. Hardware mounted above public roads may be inconvenient to reach, but it is still unattended hardware in public. The security design has to assume that someone will eventually take one.

This does not prove that Flock's cloud was breached, that a remote attacker can browse its network, or that captured cameras can poison police records. It does show that the camera itself held more recoverable data than Flock's public description led me to expect. That matters to every city buying the system and every driver passing beneath it.

## The image count changes the threat model

WIRED's analysis found that a passing vehicle typically triggered about 28 images, while some produced more than 100. The camera varied exposure, selected useful frames, cropped them, and sent images and metadata over a cellular connection. Plate reading and identification of vehicle make, model, and color appeared to happen later on Flock's servers.

The 1.6 million figure needs care. It describes images generated during the recovered log windows, not 1.6 million intact photos sitting on the camera when it was opened. Older logs had been overwritten, and much of the device's most sensitive storage remained encrypted and inaccessible. Still, the investigators recovered thousands of videos and media from a camera that Flock says should remove local data automatically after upload.

Flock's current [evidence policy](https://www.flocksafety.com/legal/flock-evidence-policy) says data stays on the camera only temporarily, is removed after upload, and remains encrypted throughout its lifecycle. Those are sensible promises. This investigation turns them into engineering questions: How quickly is local media deleted? What survives crashes? Can a key stored on the device unlock retained footage after capture?

> **Info:**
>
> Physical access is not a footnote for street hardware. It is the moment the device has to protect its keys, stored media, software, and trust relationship with the cloud.

## The camera sees more than a plate

The recovered software explicitly detected people, vehicles, license plates, and bicycles. When WIRED ran the models across 27,321 short clips from the device, they detected people in 11 clips, all riding motorcycles. The small count fits the camera's downward view over a road, where few pedestrians would appear.

The same tests found mistakes. The plate detector sometimes cropped bumper stickers, dealership frames, and other graphics as if they were plates. In one clip, it treated an American flag patch on a motorcycle saddlebag as a plate. At the same time, WIRED and 404 Media found no evidence that Flock's camera software was actively using face recognition. That distinction is worth keeping. Person detection is real; active face recognition was not found.

The local camera is only one half of the system. WIRED previously reconstructed parts of [Flock's police investigation software](https://www.wired.com/story/flock-safety-os-investigate/) and reported that server-side tools can combine camera records with police files and commercial data. The latest teardown is useful because it shows what the edge device contributes before those records reach the larger platform.

## Encryption has to survive capture

This is not the first warning about the hardware boundary. In 2025, security researcher Jon Gaines published a [technical teardown of a Flock Falcon or Sparrow reader](https://gainsec.com/2025/06/19/grounded-flight-device-2-root-shell-on-flock-safetys-falcon-sparrow-automated-license-plate-reader/). He reported missing flash encryption, disabled secure boot, and an unlocked bootloader on a unit he owned. Flock later said those findings required physical access and did not show a customer-data breach or real-world exploitation.

The new investigation does involve a deployed camera and recovered media, so the physical-access qualifier now sounds less reassuring. The hackers said an unencrypted media partition contained a key that unlocked another partition holding videos and still images. Flock told WIRED that removing and tampering with a camera is illegal, said it lacked enough detail to assess the claims, and invited submission through its vulnerability-reporting process.

That response is fair on one narrow point: taking a deployed camera is not responsible security research. A former police officer quoted by WIRED also warned that vandalism could harden official support for the system. I agree. Theft should not become the only practical route to learning what public surveillance hardware stores.

## Cities need answers they can verify

Police agencies say license plate readers help find missing people and suspects in violent crimes, a benefit documented in [Associated Press coverage of Flock's recent policy changes](https://apnews.com/article/flock-license-plate-cameras-surveillance-deflock-2a93bc075e2f7ffcca9e04a35d75a3fe). That benefit raises the standard for security rather than lowering it. Evidence used in an investigation needs a defensible chain from roadside sensor to search result.

If I were reviewing a contract, I would ask whether media keys are unique to each device, whether key material can be extracted after capture, and whether the cloud can revoke a stolen camera without trusting that camera to report honestly. I would also ask for measured deletion behavior during storage errors and reboots. WIRED found more than 27,000 "no space left on device" errors in the recovered logs, plus crashes and reboots. Failure behavior belongs in the security review.

Public reaction has already jumped ahead of the evidence. In a large [Reddit discussion](https://www.reddit.com/r/technology/comments/1whthe5/hackers_stole_flocks_camera_software_revealing/), commenters asked whether a compromised device could delete or inject records. The investigation did not establish either capability. Flock should answer that question with architecture, independent testing, and published advisories rather than leaving everyone to speculate from stolen hardware.

## My read after 21 days of logs

What surprised me was the volume created by one device, the software working beyond simple plate capture, and the amount investigators could reconstruct once the box was off its pole. A system spread across public roads cannot make physical possession an exotic scenario.

I would not read this story as instructions to pull cameras down. I would read it as a procurement test. A city should know what one captured unit can reveal, how quickly it can be cut off, what data should already be gone, and who has verified those answers. If the only clear explanation arrives after someone steals the hardware, the accountability layer is already late.