Projects

Passive 2.4 gigahertz detection for Flock Safety hardware. Runs alone, or plugs into the Flask dashboard over USB for live wardriving with GPS tags.

The cameras talk before they shoot. Every Flock unit sends out a signature on 2.4 gigahertz, and this tool just sits there and listens. Drive by and log it. That's the whole trick, and it tells you where the hardware actually lives without ever lighting up their end. Start from NitekryDPaul's original research. The dev branch layers the DeFlockJoplin information element work and a wildcard probe signature on top.

Wausau Pilot & Review's statewide tracker of license plate reader surveillance in Wisconsin. Every camera the community has mapped, every agency in the Flock data sharing network, the transparency numbers each agency posts, and hand verified contract status with sources.

Most newsrooms write a story about the cameras and move on to the next thing. These people kept the receipts. Every Monday the ledger pulls the public records again and reposts the map and the numbers. Any small outlet can run the same playbook in its own state, and that's the point.

99,435,526 deduplicated ALPR search records from 6,466 law enforcement agencies, 2023 through 2026. Every record released by a government agency under public records law.

Every row in this pile came out of a public records request. Nobody scraped a system, nobody bought a database. The sources.json file ties each chunk back to the exact request that freed it, so the whole thing is auditable if you ever doubt it. A hundred million rows is enough to finally see whose plates get searched and how often.

The first knowledge database corruption attack against retrieval augmented generation, published at USENIX Security 2025.

Copilots answer from a pile of stored documents. Get your hand into that pile, and a few poisoned pages bend the answer the way you want. Agencies are already strapping these things onto plate reads and records searches. Read the paper and see how cheap the corruption really is.

The standard Python library for machine learning security. Evasion, poisoning, extraction, and inference attacks, with defenses to match. Built for red teams and blue teams.

The camera companies sell their models as airtight, but a model is just a machine, and machines can be fooled. ART is the toolbox for testing these models the way an adversary would. Run the attack before somebody else runs it on you, then show the public what falls apart.