ICE seeks AI platform to analyze billions of investigative records

U.S. Immigration and Customs Enforcement (ICE) is exploring a centralized investigative platform that would place generative AI on top of billions of communications, location, financial, social media and digital forensic records, allowing Homeland Security Investigations agents to query and analyze massive datasets using ordinary language.

The proposed capability would allow investigators to rapidly generate timelines, maps, relationship graphs and other analytical outputs from information drawn across multiple surveillance and investigative systems.

Judging by the RFI, ICE is not looking for another tool to store wiretap records or plot cellphone locations, but rather an environment in which communications, movements, social media activity, financial information and digital evidence can be brought together at enormous scale and probed conversationally using AI.

The system ICE is contemplating is revealed in the innocuous sounding August 10 Request for Information (RFI) for a “Telecommunications Analysis System.” ICE describes it as a criminal investigative platform, not a mass-surveillance tool.

However, the capability ICE is exploring illustrates some of the same questions about AI and government surveillance that erupted earlier this year in the Pentagon’s confrontation with Anthropic over limits on the use of the company’s Claude AI models.

At the center of that dispute was Secretary of Defense Pete Hegseth’s position that AI companies supplying the military should permit their technology to be used for any lawful purpose. Anthropic agreed to broad national security uses but refused to remove restrictions in its terms of service covering fully autonomous weapons and what it calls “mass domestic surveillance.”

The Pentagon said it had no intention of conducting mass surveillance of Americans, but that U.S. law, rather than a private company, should determine what is a permissible government use.

Hegseth led the dispute with Anthropic, at one point complained about “social justice infusions that constrain and confuse our employment of this technology.”

The “Department of War AI will not be woke. It will work for us,” he said while arguing military AI must operate without what he called ideological constraints. He added, “We’re building war-ready weapons and systems, not chatbots for an Ivy League faculty lounge.”

There was good reason for Anthropic to be concerned.

Anthropic said it supports AI for lawful foreign intelligence and counterintelligence purposes, but argues that AI creates new privacy risks because it can automatically assemble scattered information about people’s movements, web browsing and associations into comprehensive dossiers on their lives at massive scale.

Anthropic’s explanation for drawing a line on surveillance is particularly relevant to the system ICE is considering.

ICE wants a centralized AI-powered telecommunications analysis platform that can support thousands of HSI users, including at least 500 concurrent users spread across U.S. and foreign offices, which can process, search, store and analyze datasets of at least 10 terabytes encompassing billions of communications records, while also absorbing ICE’s legacy database files and active case data.

ICE said the RFI is market research, meaning it isn’t seeking bids or committing itself to buy the system it describes in the RFI. Responses are due August 18.

The range of information the proposed system would handle would certainly be massive.

ICE wants the platform to be able to receive and analyze several kinds of live, lawfully authorized surveillance data in one place, including Title III wiretaps and other live call interceptions, pen-register and trap-and-trace data associated with telephone, email, social media, IP addresses and WhatsApp communications, as well as live GPS and location pings.

Historical information would include subscriber records, call detail records, cell tower information including tower dumps and timing-advance data, GPS locations, cyber tips, web browsing histories, email, financial information and transactions and records from services including Google, Microsoft, Snapchat, Instagram, Facebook, X and WhatsApp.

Evidence extracted from mobile devices would be another part of the same analytical environment. ICE specifically identified data obtained through Magnet Axiom, Cellebrite, Oxygen Forensics, Susteen and XRY.

Magnet Forensics is a comprehensive digital forensics and incident response platform used to recover, analyze, and report evidence from mobile devices, computers, cloud services, and vehicles.

Cellebrite provides end-to-end digital forensics, investigations, and intelligence solutions.

Oxygen Forensics provides digital forensics software to extract, decode, and analyze data from mobile phones, cloud services, drones, and IoT devices – even when that information is locked, encrypted, or deleted.

Susteen provides digital mobile forensics, data recovery, and investigative tools, as does XRY.

The capability ICE is exploring would allow its HSI investigators to correlate all of this type of information through searches, filters, dashboards, graphs and timelines, as well as geo-temporal mapping of call records, mobile forensic data, GPS locations, other location-tracking information and social media records.

ICE said the system would also need the capabilities for 3D mapping, mapping IP addresses, live call alerts and geofence alerts for GPS locations. An integrated phone application would also make live information from the system available to mobile users and be able to look across active investigations.

And on top of all of this data, ICE would embed a generative-AI assistant operating with ICE-provided Azure AI accounts.

ICE Azure AI accounts is significant because HSI is not asking the telecommunications system contractor to supply a separate, self-contained AI service. It wants the vendor’s surveillance and analytical platform connected to AI infrastructure ICE already controls within Microsoft’s cloud environment.

ICE’s reliance on Azure has been expanding rapidly. The amount of data it stores in its Microsoft Azure cloud has more than tripled.

According to the RFI, the contractor would provide the analysis system but the generative-AI assistant would operate through ICE/HSI-provided Azure accounts, an arrangement that would keep the AI layer inside ICE while giving the agency greater control over the models and services used to interrogate its investigative data.

In practical terms, ICE is looking for a vendor-built surveillance platform that it wants wired directly into an AI infrastructure it already controls, rather than buying the two as a single proprietary package.

Instead of requiring an investigator to construct a series of conventional database searches, ICE says users should be able to use natural language prompts to “search, query, filter, interact with, summarize, map, and visualize” complex datasets.

ICE also wants the system to be able to produce lists, maps, timelines, tables, graphs and charts and have interfaces for AI-assisted image tagging, translation and transcription.

Investigative data, user prompts and resulting outputs would not be used to train external public or commercial AI models, the RFI says, and every AI interaction would have to be logged so it can be audited and traced.

Anticipating the future, ICE said what it is looking for must accommodate current and future formats required under the Communications Assistance for Law Enforcement Act (CALEA), the 1994 federal law requiring telecommunications carriers and certain broadband and communications providers to design and operate their networks so they can carry out lawfully authorized electronic surveillance.

Article Topics

generative AI  |  ICE - U.S. Immigration and Customs Enforcement  |  procurement  |  surveillance  |  U.S. Government