Hackers pulled down a Flock camera and found the encryption key inside

Why it matters

Flock's value depends on cheap cameras surviving in public while preserving the evidence inside. A recovered key turns physical tampering into a product-security and customer-trust problem.

Hackers physically removed a Flock Safety camera from above a roadway, copied much of its internal storage and recovered an encryption key that opened videos and still images captured by the device. The hardware compromise, detailed in a joint WIRED and 404 Media investigation and 404 Media's account published September 16th, shows how a roadside camera in Flock's national surveillance network processes vehicles and people before sending evidence to the cloud.

Langley founded Flock Safety in 2017 with Matt Feury and Paige Todd after break-ins in his Atlanta neighborhood left police without useful evidence. The founders' first version was effectively an Android phone in a waterproof box, Todd recalled in a Georgia Tech profile. Cheap components, cellular connectivity and machine learning helped Flock replace automatic license plate readers that could cost about $25,000 per camera.

That origin now sits at the center of Flock's security test. Langley built Flock by making connected cameras inexpensive enough to mount across neighborhoods and police jurisdictions. Every device placed above a public road also becomes a computer operating beyond the physical control of Flock's engineers and customers.

The key was stored with the evidence

The hacker collective stegan0gram said it accessed the camera's Android system and found unencrypted partitions identified as "vendor" and "media." The media partition held a key that unlocked another storage area containing much of the device's captured video and imagery.

The compromise did not expose everything. WIRED and 404 Media reported that some of the camera's most sensitive storage remained encrypted and inaccessible. The recovered material was still extensive: logs covering roughly 21 days showed about 50,200 vehicle detections and 1.6 million images. A typical passing vehicle generated approximately 28 images, while some produced more than 100.

That observed behavior differs from Flock's public explanation of its cameras. Flock's frequently asked questions say its license plate readers generally capture six to 12 images of a vehicle per session, with engineers able to adjust the settings for traffic and location. The compromised camera averaged well above that range during the periods covered by its surviving logs. The findings establish the operation of this particular device rather than every camera model and configuration in Flock's network.

Flock also says images and metadata are encrypted throughout their lifecycle, beginning on the device and continuing through transmission and cloud storage. That claim can remain accurate at the protocol level while offering limited protection when a decryption key is recoverable from the same physical hardware. An attacker does not need to defeat the encryption algorithm if the deployed device supplies the key.

A Flock spokesperson told WIRED and 404 Media that removing and tampering with a camera is illegal. Asked about the recovered key, Flock pointed to its public vulnerability disclosure policy and said security researchers should report potential flaws directly. The response leaves the engineering issue exposed by this camera intact: physical possession provided access to stored media that Flock had described as protected.

The camera was looking for people too

The recovered software showed that the device explicitly detects people, bicycles, vehicles and license plates. Researchers extracted the computer-vision models and ran them against test images and 27,321 short video clips recovered from the camera.

The models detected people in 11 clips, all involving motorcycle riders. The low count reflected the camera's position above a road, where pedestrians rarely entered its field of view. WIRED and 404 Media found no active facial-recognition system beyond capabilities included by default in Android.

The software also treated some bumper stickers, dealership frames and other graphics as possible plates. In one case, it cropped an American flag patch on a motorcyclist's saddlebag. Those errors reveal how the edge software reduces a stream of roadway activity into selected objects and images before uploading results over a cellular connection.

The camera itself did not appear to read plate text or identify vehicle makes, models and colors. Those functions appeared to run on Flock's servers. The roadside hardware handled motion detection, rapid image capture, exposure changes, object classification, cropping, uploads and remote updates through about 20 Flock-built applications.

That division of work matters because Flock's network is commonly discussed as a searchable cloud database. The compromised device shows how much processing and temporary evidence collection happens at the edge, where physical defenses become part of the privacy model.

Langley's trust problem has reached the hardware

Langley has consistently framed Flock as an attempt to replace subjective policing with objective evidence. In August, Flock published his argument that safety should be treated as a right and said its systems supported over 1 million investigations and incidents during 2025. Those figures remain Flock's claims, but the message captures the founder's central bet: broad camera coverage can reduce crime while policy controls, short retention periods and audit logs limit abuse.

The latest compromise adds device security to a trust fight already playing out over access and permissible searches. RuntimeWire previously reported that more than 2,000 organizations could access plate data from one Atlanta suburb, while Boston ended its Flock pilot after records were shared outside the city despite a contractual restriction.

Flock's software ambitions have also moved beyond matching plates. RuntimeWire reconstructed how OS Investigate can connect vehicle records with behavioral searches, and a later examination found that AI watchlists can continuously search for people matching written descriptions. The camera teardown supplies the other half of that system: the code and sensors that decide which pieces of the physical world become searchable evidence.

Investors have financed Langley's expansion at a scale rarely available to public-safety hardware startups. Flock raised $275 million at a reported $7.5 billion valuation in March 2025, with participation from Andreessen Horowitz, Greenoaks, Meritech Capital, Matrix Partners and Y Combinator. TechCrunch reported at the time that Flock had surpassed $300 million in annual recurring revenue during 2024, according to a Flock spokesperson.

That money has helped Flock grow from neighborhood plate readers into cameras, gunshot detection, drones and investigation software. It also raises the standard Langley must meet. A public roadside computer should be designed on the assumption that determined people will eventually reach it, remove it and inspect every component.

Stegan0gram said it intends to publish details that could help others repeat the extraction. Wider replication would turn a single teardown into a direct test of Flock's deployed fleet. For Langley, the work now extends beyond convincing cities that cameras solve crimes and access policies prevent misuse. Flock has to prove that the evidence inside each camera remains protected when the camera itself falls into someone else's hands.