1984: American Surveillance System

"'Never Hacked': The National Security and Accountability Failures in Flock Safety's Nationwide Dynamic Surveillance Network" is Joshua Michael's (Nexanet.ai) independent cybersecurity research package: a 335,701-record production camera export, four reported vulnerabilities, and modeled national-security exposure studies at 22 sensitive sites. It's licensed CC BY 4.0 — anyone may use it, including commercially, with credit to him.

We overlay this inventory, unmodified apart from coordinate rounding and dropped empty columns, on Sound Mapper Pro, alongside our own RF coverage modeling for each device class. We don't mirror his full evidence package or the national-security studies here — that's his work, published at his discretion, and we point to it rather than rehost it.

Changes we make to the data when we display it: coordinates rounded to the nearest few meters, a handful of invalid rows (lat/lon both zero) dropped, and five already-empty identifier columns removed. Everything else, including device status and type, is shown as published.

Across these logs, covering roughly mid-2024 through mid-2025, we count 4,641 distinctly-named outside agencies reaching into these camera networks. The heaviest users: Houston TX PD (276,439 searches), Gwinnett County GA PD (248,885), Dallas TX PD (209,408), Atlanta GA PD (127,891), and Cobb County GA PD (122,336) — all querying into departments hundreds to over a thousand miles from their own jurisdiction.

Methodology, read before citing this: we parsed the raw exports with every row's Name, License Plate, Case #, and Filters fields stripped out before any count was taken — those fields never left the aggregation step. "Reason" categories above come from simple keyword matching on department-entered free text (e.g. a reason containing "stolen" counts as stolen-vehicle); it's a rough heuristic, not a verified classification, and the single largest bucket is text that didn't match anything we looked for. Duplicate files (byte-identical exports saved under two names) were counted once. Agency names are taken as departments typed them, so the same department under slightly different spellings may be counted more than once in the 4,641 figure.

This is a first pass, not the full picture. Beyond these three departments, we have roughly 14.7 million additional rows of nationwide network-audit exports still to deduplicate and reconcile against the departments above — each capped at Excel's 1,048,576-row export limit, so even that is a truncated slice. Separately, we're sitting on close to 25 GB of further Flock audit-log exports from police departments across the country that haven't been ingested yet. The numbers here will grow; they won't shrink.

The current wave of data-center construction is sold to the public as generic AI infrastructure — training and inference capacity for chatbots and enterprise software. That framing is accurate but incomplete. The same inference capacity is also the compute backbone for the AI-enabled surveillance systems being installed alongside it: automated plate reading, object/vehicle classification, and cross-network search at the scale the audit logs above show. A rezoning hearing about "an AI data center" rarely mentions that connection.

This is our working thesis, stated plainly as analysis rather than a sourced claim — we're holding off on publishing specific figures (grid load, tax cost, local impact) until we've assembled proper citations for them. Sound Mapper Pro's RF and device-coverage modeling is the tool we built to make the capability side of this concrete; the infrastructure side is still in progress.