Police hugging AI so tight, soon we’ll all be pre-crime suspects

The risks presented by superintelligent and uncontrollable artificial intellifence are real, but the security appartatus is already colloborating closely with these firms

Harrison Berger – Responsible Statecraft:

Earlier this month, an announcement by an Anthropic researcher that so-called “Superintelligent AI” has a greater than 10% chance of destroying humanity over the next decade triggered widespread alarm. “We really do earnestly believe AI could kill all humans,” wrote Evan Hubinger, Anthropic’s Alignment Science Lead.

Those risks are, of course, real and terrifying. But perhaps of more immediate concern are the ways in which humans are already using AI, not how AI will potentially escape human control.

Though it has been mostly ignored by corporate media, the U.S. security state is now seeking out and collaborating with AI firms on “offensive cyber ops”; AI targeting systems; and efforts to sift through and interpret vast troves of data on American citizens and foreigners collected by the U.S. government.

A dispute between the San Francisco-based company Anthropic and the Department of Defense — followed by reporting that its AI model, Claude, plays a role in the targeting system used in the U.S. bombing of Iran — brought the relationship between AI firms and the U.S. security state into mass public consciousness.

Anthropic’s refusal to sign an “all lawful uses” contract with the Pentagon and its subsequent designation as a “supply-chain risk” garnered the company widespread praise, producing what AP called “a surge of Claude downloads that made it the most popular iPhone app.” The company, which describes itself as committed to ethical limits on AI, saw its valuation skyrocket to $965 billion in its latest funding round.

But Anthropic’s high-profile dispute obscured how AI firms, including Anthropic, are already deeply integrated into the U.S. military-industrial complex.

A confidential working paper leaked to tech journalist Jack Poulson, titled “The Art of the Possible,” documents a closed-door workshop in November 2023 at which Anthropic representatives met with senior CIA officials alongside officials from Australia’s Office of National Intelligence. The event was devoted to how large language models could be integrated into Western security agencies.

As Poulson reported, Anthropic has since hired the longtime Palantir employee Steve Sloss to lead its pitch to U.S. intelligence agencies, including for military targeting, and partnered with Palantir “to provide U.S. intelligence and defense agencies access to the Claude 3 and 3.5 family of models on AWS [Amazon Web Services].”

Inside Palantir’s Maven system, military planners have used Claude to analyze intelligence feeds and build prioritized target lists for strikes on Iran. Claude also played a role in the operation to kidnap Venezuela’s leader Nicolas Maduro, according to the Wall Street Journal. Asked by Bloomberg about the possibility Claude was used in the strike on a girls’ school in Iran that killed at least 120 children, Anthropic CEO Dario Amodei said that he did “not know exactly” if or how Claude had been used in the strike but that the girls school bombing was “a use case that doesn’t even violate our red lines.”

And, according to Financial Times, despite its supply-chain risk designation, Anthropic helps the NSA “deploy its powerful Mythos AI model for offensive cyber operations, embedding engineers inside the agency.” One person close to Anthropic defended the cyber ops, telling FT that “if [Mythos] is not used to build attack agents, adversaries will find a way to do it.”

OpenAI’s contracts with the U.S. military and security state are equally significant, yet the firm’s largest public payout from the federal government had until recently drawn little attention. On October 16, 2025, the company signed a one-year, $3.5 million Other Transaction Agreement with U.S. Cyber Command for “AI platform training,” as Poulson first reported. Cyber Command was among the first participants in the Iran War, conducting preparatory cyberattacks against Iranian communications infrastructure. Their contract was signed 16 months after OpenAI announced the appointment of Paul Nakasone — who had led Cyber Command and the NSA jointly until his retirement in February 2024 — to its board of directors.

AI surveillance
Having too much data — and not knowing how to interpret all of it — was one of the main problems for the NSA when they introduced their Trailblazer bulk collection program. Steven Cash, an architect of the Patriot Act, testified to Congress in June that “Artificial intelligence and machine-learning systems now permit forms of analysis [of bulk collected data] that would have been unimaginable even a decade ago.” Procurement documents and independent media reports demonstrate that interpreting large sets of data on Americans and foreigners is exactly how the U.S. security state intends to use its new AI tools.

In March, The New York Times revealed that the Pentagon’s chief technology officer, Emil Michael, had requested Anthropic “allow for the collection and analysis of unclassified, commercial bulk data on Americans, such as geolocation and web browsing data.” That data, known as commercially available information (CAI), has been increasingly purchased and hoarded by the security state, as the Director of National Intelligence acknowledged in a 2023 report. “CAI,” the report warned, “can reveal sensitive and intimate information” and “be misused to pry into private lives, ruin reputations, and cause emotional distress and threaten the safety of individuals. Even subject to appropriate controls, CAI can increase the power of the government’s ability to peer into private lives to levels that may exceed our constitutional traditions or other social expectations.”

Meanwhile, a publicly viewable request on the federal government’s procurement website SAM.gov seeks AI tools for what journalists like Dan Boguslaw describe as “pre-crime policing” to be deployed at the FBI’s Threat Screening Center (TSC), with “Predictive Modeling” as one of the six requirements for potential AI vendors.

The TSC, which maintains the federal terrorist watchlist created after 9/11, was renamed in 2025 from the “Terrorist Screening Center” to the “Threat Screening Center,” according to the agency website, “to reflect an expanded mission” which goes “beyond terrorism watchlisting and screening to address other national security threats.”

Reporting by Ken Klippenstein indicates TSC is expected to administer the new category of domestic threats created by NSPM-7, the controversial authority that calls for a “comprehensive” strategy to surveil, investigate, and prosecute what it labels “domestic terrorism and organized political violence,” identifying “anti-Christian,” “anti-American,” “anti-capitalist” and “anti-fascist” views as indicators of radical-left violence. “The TSC,” writes Boguslaw, “seeks to prosecute crimes before they occur,” while a 2026 report by the Government Accountability Office indicates TSC “shares surveillance data with state and local law enforcement.”

Those tools would join a growing series of AI integrations already operating across the U.S. security state. In September 2025, ICE’s Homeland Security Investigations unit signed a $5.7 million contract through Zignal Labs, a social media monitoring platform. ICE has meanwhile paid Palantir $30 million to build ImmigrationOS, while 404 Media separately reported that Palantir is creating a tool for ICE that “populates a map with potential deportation targets, brings up a dossier on each person, and provides a ‘confidence score’ on the person’s current address.”

Other tools include Babel X, the AI-enabled Babel Street product used by the Department of Homeland Security to gather data on various groups, including U.S. citizens.

The DHS’ AI surveillance apparatus is set to expand. Contract forecast documents reviewed by FedScoop show the agency planning to spend well over $100 million on various AI-powered surveillance systems.

Tech researcher Jack Poulson has argued that commercial large language model usage on platforms like ChatGPT and Claude will become central to police and intelligence interception, citing the case of Nebraska-based wiretapping firm PenLink, which has trained its law enforcement customers on gaining access to the Google search histories of investigative targets.

While the risks presented by superintelligent and uncontrollable AI are real, the current debate misses the ways in which AI is already being deployed by the security state and military industrial complex against foreign and domestic populations.