Loading story…
Cybersecurity researcher Bill Swearingen publicly demonstrated Project noRecognition, a machine-learning system that generates printed wrap patterns intended to make vehicles or people harder for AI surveillance systems to detect. At DEF CON, he showed a wrapped 2009 Toyota Yaris passing a live Flock Safety camera; the camera reportedly captured video but did not log the vehicle through its automated detection pipeline. Swearingen said the system was trained with reinforcement learning across roughly 31 million tests and evaluated against 11 open-source detection algorithms, including software linked to license plate readers, body cameras, and facial recognition workflows.
The project aims to create privacy-preserving patterns that defeat automated tracking without simply obscuring a license plate, but reported results remain mixed. Security reporting said the strongest validated outcome on the project dashboard was 61.7% non-detection against a detector derived from a real deployed surveillance camera, and that result was achieved in digital simulation rather than consistent field conditions. Swearingen said real-world testing showed promise, though practical limitations remain, including weak points such as exposed wheels and the broader gap between simulated evasion and reliable operational use.
See the reporting duties and controls this puts on the clock.
5 events from the most recent confirmed update back to the earliest known activity.
Swearingen publicly demonstrated that a live Flock camera recorded video of the wrapped vehicle but reportedly failed to log it through its automated detection software. He presented noRecognition as a privacy-preserving technique intended to let people opt out of automated tracking without obscuring license plates.
At DEF CON in Las Vegas, Swearingen carried out his first real-world test by wrapping a 2009 Toyota Yaris in a noRecognition pattern and driving it past a Flock Safety camera. He said the test was effective, though the vehicle's wheels remained a weak point for evasion.
The noRecognition project dashboard cited its strongest validated result as 61.7% non-detection against a detector extracted from a real deployed surveillance camera. The article notes this result was still obtained in digital simulation rather than a real-world field test.
Swearingen said he spent about a year running roughly 31 million tests to train noRecognition to generate printed patterns that cause surveillance AI systems to miss covered vehicles or people. The system iteratively tested patterns against detection software and adjusted them when they were detected.
Bill Swearingen said he started the noRecognition project because he was concerned about being tracked by surveillance cameras while attending a protest. The effort grew from an initial experiment into a reinforcement-learning-based system for generating detection-evasion patterns.
Vulnerabilities, threat actors, malware, products, organizations, and breaches Mallory has linked to this story.
See what this changes for your reporting obligations and which controls it puts on the clock.
2 references tracked. Mallory keeps watching after this page renders.
Map indicators from this story to your assets and identify affected systems in minutes.
Every observed campaign, victim, and pivot linked to actors named in this story.
Malware, exploits, and IOCs connected to the activity described here.
YARA, Sigma, and Snort rules deployed to your SIEM as soon as they’re published.
Get matching new stories delivered to your team as they break — not the next morning.
Ask questions about this story and take action on the answers.
Other stories Mallory is tracking that overlap on entities, sources, or topic labels.
Researcher Bill Swearingen publicly unveiled **noRecognition**, a project that uses reinforcement learning to generate adversarial visual patterns designed to stop automated surveillance systems from recognizing people, objects, and vehicles while still allowing ordinary video recording. Swearingen said the system ran roughly **31 million tests** over the past year and produced increasingly effective patterns by learning from prior detection failures. In lab testing, Swearingen said the patterns defeated **11 open-source detection algorithms**, including software associated with **Flock** license plate readers, **Axon** body-worn camera workflows, and **Clearview AI**-enabled camera pipelines. He also demonstrated the concept at **Def Con** by covering a **2009 Toyota Yaris** with one of the patterns and testing it against a Flock camera, while keeping the most effective designs offline to slow vendor countermeasures and pursuing crowdfunding for derivative products such as shirts, hoodies, and possible vehicle wraps.
Aug 11, 2026
Flock Safety’s network of roughly **90,000** vehicle-surveillance cameras is facing renewed scrutiny after a false alert tied to partial license plate matching reportedly led Minnesota police to stop and arrest a driver in a Jaguar Land Rover media fleet vehicle. According to the account, a stolen New Jersey manufacturer plate was entered into Flock’s system as `34 DTM`, omitting smaller middle digits and causing the platform to match a broader `34 ## DTM` pattern. Flock said the system operated as configured and that law enforcement wanted partial-plate matching so officers could receive alerts even when plate data is incomplete. The incident has intensified criticism of Flock’s broader surveillance model, which uses Falcon automatic license plate reader cameras and Condor pan-tilt-zoom cameras to capture vehicle images and metadata, transmit them over cellular networks, and store searchable records in Flock’s cloud platform on Amazon Web Services. Privacy advocates and former law enforcement officials warn that shared cross-agency access, AI-assisted searches, insider misuse, and person-tracking capabilities can reconstruct individuals’ movements and turn bad data into dangerous police encounters. Public opposition has also grown in some communities, including Saranac Lake, where residents pushed officials to cancel a Flock contract and remove cameras.
Jul 21, 2026
The hacker collective **stegan0gram** said it removed a Flock Safety automated license-plate-reader camera from a roadway, disarmed it, and reverse engineered its hardware, solar equipment, and Android-based operating system. The group copied nearly all data held on the device—including thousands of videos, still images, and operational logs—and shared material with WIRED and 404 Media. The extracted data reportedly showed one camera captured roughly **1.6 million images of 50,000 vehicles in 21 days**, illustrating the scale of Flock's vehicle-tracking collection. Researchers said unencrypted `vendor` and `media` partitions were accessible; an encryption key in the media partition allegedly unlocked another partition containing much of the captured imagery. Flock cameras upload vehicle images and associated metadata to searchable Flock servers used by authorized agencies, including participants in its national network.
Sep 23, 2026