LIVE FEED
FIRST LOOK ATLAS OWASP MEDIUM Moderate risk · Monitor closely RELEVANCE ▲ 5.5

Flock Safety Launches OS Investigate AI for Law Enforcement

FIRST LOOK MEDIUM ↗ MODERATE
  • What shipped: Flock Safety's OS Investigate ships AI-powered multi-source investigative queries for law enforcement across 6,000+ communities.
  • Who benefits: Law enforcement agencies and their oversight bodies benefit most — the tool closes manual correlation gaps but requires governance frameworks to ensure lawful, auditable use.
  • Next steps: Establish documented access-control and authorisation policies before deploying OS Investigate in operational environments. · Audit the 45 integrated tool connections and data-source permissions to ensure minimum-necessary access is enforced. · Implement query logging and supervisory review workflows for all natural-language prompts submitted to the system.
Flock Safety Launches OS Investigate AI for Law Enforcement

Defender Impact

Flock Safety’s OS Investigate closes a longstanding gap in law enforcement intelligence fusion: the inability to correlate licence-plate camera data, case files, dispatch logs, and commercial identity records through a single, accessible interface. For investigative teams, the shift from manual cross-database triage to natural-language agentic queries represents a genuine capability uplift — but realising the full benefit requires governance maturity that the current deployment phase has not yet demonstrated publicly.

Capability Overview

OS Investigate (previously codenamed Nightshift) is an AI-driven investigative platform built on top of Flock Safety’s existing Automated Licence Plate Reader (ALPR) network, which logs vehicle movements across more than 6,000 communities. The system exposes 45 integrated tools that the AI can invoke autonomously in response to officer queries, spanning plate scans and camera metadata, arrest records, case files, 911 dispatch logs, ballistics results, and commercial databases containing Social Security numbers, dates of birth, phone numbers, email addresses, and associate graphs.

The interface ships with 69 preloaded natural-language prompts that officers can select, edit, and submit via a chat interface. Officers may also author custom prompts. The architecture is explicitly agentic: the LLM selects and chains tool calls based on the prompt, returning synthesised results rather than raw database records. Crucially, the system inverts the traditional ALPR model — where a known plate is checked against a watchlist — enabling population-down queries where an officer supplies a location, time window, and behavioural pattern, and the system identifies matching individuals.

WIRED recovered over 450 cached files from Flock’s own web infrastructure, including the prompt library and tool definitions, confirming the system’s scope ahead of any formal public launch. Flock describes the product as still in limited testing with a small number of law enforcement partners.

Defensive Advances

Cross-source intelligence fusion at operational speed. Investigators can now issue a single natural-language query and receive correlated results from camera networks, case management systems, and commercial identity data — a task that previously required analyst hours and manual API access across disparate systems.

Pattern-of-life witness identification. The ability to surface individuals whose vehicles repeatedly pass through a crime scene area provides investigators with a structured, repeatable methodology for identifying potential witnesses, replacing ad hoc manual review of raw ALPR logs.

Reduced skills barrier for structured data analysis. Preloaded prompts and a chat interface mean that complex investigative queries — previously requiring database expertise — are accessible to frontline officers, broadening the effective user base for structured intelligence analysis.

Geospatial population search. The map-draw search function enables rapid geographic scoping of investigations, a capability that previously required dedicated GIS analyst support.

Residual Gaps

The current deployment phase leaves several maturity questions unanswered. First, query audit trails: it is not publicly documented whether all natural-language prompts, tool invocations, and returned results are logged in a tamper-evident, legally reviewable format — a prerequisite for defensible use in judicial proceedings. Second, access-control granularity: with 45 tool connections spanning sensitive commercial databases, the absence of publicly documented role-based access controls and minimum-necessary data principles is a significant operational gap. Third, prompt governance: the preloaded prompt library was recoverable from Flock’s public web infrastructure, raising questions about whether the system’s query surface is adequately bounded in production environments. Fourth, oversight integration: no public documentation confirms integration with existing legal-hold, warrant-tracking, or supervisory approval workflows — a requirement for constitutionally compliant deployment in most jurisdictions.

Framework Mapping

  • LLM08 (Excessive Agency): The system autonomously selects and chains 45 tool calls based on officer prompts, creating an agentic surface where unintended tool invocations could return data beyond investigative scope.
  • LLM06 (Sensitive Information Disclosure): Aggregation of SSNs, DOBs, and associate graphs into LLM-synthesised outputs raises output-handling risk if responses are inadequately scoped.
  • LLM09 (Overreliance): Preloaded prompts and natural-language interfaces may create investigative overreliance on AI-synthesised results without independent source verification.
  • AML.T0086 (Exfiltration via AI Agent Tool Invocation): The agentic tool-chaining model is a surface that warrants monitoring for unintended data aggregation across tool boundaries.

Deployment Considerations

Organisations evaluating OS Investigate should treat governance infrastructure as a prerequisite, not a follow-on. Before operational deployment: establish documented authorisation policies mapping query types to legal authorities; configure role-based access controls limiting which tool connections are available to which officer roles; and integrate query logging into existing evidence management and legal-hold systems. Agencies should also conduct a data-minimisation review of the 45 tool connections to confirm that each integration has a documented legal basis and defined retention limit.

Defender Checklist

  • Document the legal authority (statute, warrant type, policy basis) required for each of the 45 tool-connection categories before enabling them in production.
  • Implement tamper-evident query logging covering prompt text, tool calls invoked, and results returned for every OS Investigate session.
  • Establish supervisory review workflows for high-sensitivity query types (geospatial population search, associate graph queries).
  • Conduct a data-minimisation audit: disable commercial database tool connections that are not required for the agency’s primary investigative mandate.
  • Test the prompt interface against internal red-team scenarios to identify unintended data aggregation paths before broad officer rollout.
  • Align OS Investigate deployment with existing body-worn camera and evidence management audit standards to ensure investigative continuity.

References

◉ AI THREAT BRIEFING

Stay ahead of the threat.

Twice-weekly digest of critical AI security developments — every story mapped to MITRE ATLAS and OWASP LLM Top 10. Free.

No spam. Unsubscribe anytime.