Flock Safety has built an AI system that moves far beyond checking whether a passing license plate belongs to a wanted vehicle. The police technology, first called Nightshift and now named OS Investigate, can analyse movement patterns, identify possible witnesses, infer associations between drivers and connect vehicle records with names, addresses, relatives and other personal information.
WIRED reconstructed parts of the system from more than 450 files served publicly by Flock’s own login pages. The reporting found 69 prepared prompts and descriptions of 45 tools reaching across camera metadata, police records, dispatch logs, ballistics results and commercial identity databases. Flock says the product remains in development and is being tested with a small group of law-enforcement partners.
Investigations can begin with a pattern instead of a suspect
Traditional plate-reader use starts with a known vehicle. OS Investigate reverses that sequence. An officer can supply an area, a period and a type of behaviour, then ask the system to produce the people or vehicles matching it.
Prepared searches include finding cars repeatedly seen in a neighbourhood as potential witnesses, identifying vehicles that visited several banks or shops, and tracing trips between cities. Nineteen prompts focus on patterns; 14 require no plate, name or physical description. Another workflow can begin with people arrested more than twice for almost any offence, map their homes, retrieve service calls and create detailed profiles of selected individuals.
The system can also infer “associates” of a target vehicle. Code reviewed by WIRED counts cars appearing at the same cameras within a two-minute window and can return up to 20 vehicles when they cross a default frequency and confidence threshold. Proximity becomes a machine-generated relationship, even though travelling near another car does not establish a personal connection or criminal involvement.
Accessible code reveals scope, but not every server-side control
The exposed files allowed WIRED to reconstruct interface elements, prompts, search fields and tools without accessing authenticated police data. Independent researcher Buchodi reproduced key parts of the analysis. The code shows fixed descriptions of data sources and workflows, but it does not reveal the model’s hidden instructions, server responses or whether certain requests are refused.
Before a search, the interface can require a written justification. The visible code appeared to impose no meaningful standard on that explanation, while an optional case number needed only three characters. WIRED could not determine whether Flock performs additional validation after submission.
Flock did not dispute the reported capabilities. A spokesperson described OS Investigate as separate from the company’s plate-reader product and said it helps agencies work across information they already possess. The company also cautioned that functions may change before wider release. That qualification matters, but it does not remove the significance of the prompts and tool access embedded in the tested software.
Weak oversight could turn convenience into mass suspicion
Privacy and legal specialists told WIRED that pattern-led searching risks treating ordinary activity as a basis for suspicion. A physical-description search can combine traits such as race, height, clothing, scars or tattoos with a radius or a shape drawn on a map. Other prompts could flag routine travel simply because it resembles a behaviour selected by an officer.
The concern is reinforced by documented misuse of existing plate-reader systems. A Washington Post analysis cited by WIRED found 50 officers charged with or accused of misuse, including 26 cases involving women they knew or wanted to meet. An Electronic Frontier Foundation review of more than 12 million searches found roughly 20% used vague justifications, while more than 50 agencies recorded hundreds of searches referring to protests.
Flock says new safeguards will monitor abnormal searches, automatically lock flagged users and require case codes by year-end. Yet many governance choices remain with individual departments, including credentials, data sharing and optional audit settings.
OS Investigate demonstrates how natural-language AI can transform a large surveillance archive from a lookup tool into an engine for open-ended inference. That may accelerate legitimate investigations, but it also lowers the effort required to search across people who were never identified as suspects. The central test is not merely whether the AI finds useful patterns. It is whether law, oversight and technical controls can stop those patterns from becoming automated grounds for suspicion.
