The Trump administration has shifted immigration enforcement from the "worst of the worst" rhetoric of the campaign trail to a quieter, data-driven strategy that uses AI to find and arrest the people easiest to catch - many of them lawfully present in the United States.
More than 6 million immigrants with temporary protections are potentially exposed, including Temporary Protected Status holders, DACA recipients, and people with pending asylum cases or humanitarian parole. According to the American Immigration Council, immigration-related AI use cases at DHS rose 36% by January 2026 over 2024 levels. Internal ICE documents describe "target-rich" environments - immigrant enclaves where arrests will yield the largest haul of detainees.
How DHS builds its surveillance web
DHS gains access, through questionably legal means, to confidential personal data from federal sources including the IRS, state Medicaid rolls, and SNAP records. It then layers that onto its own administrative records from immigrants' U.S. Citizenship and Immigration Services (USCIS) filings.
Raw data isn't especially useful on its own. The bulk of DHS spending on immigrant surveillance goes to aggregating that data to show how to locate millions of potential targets through analyses of patterns in individual behavior. While DHS spent $228.6 million on data purchase contracts since January 2021, it has spent $622.8 million on data analytics. The enforcement strategy still relies heavily on human intelligence from 287(g) partnerships with local law enforcement, but it's increasingly enhanced by evolving surveillance tools.
The guardrails are coming off
Last July, a test version of OpenAI's ChatGPT software, running without its usual safety guardrails as part of a worst-case capability test, used a previously unknown vulnerability to access the internet and breach the infrastructure of AI startup platform Hugging Face. Hugging Face's chief security officer noted that over five days, the company's agent took 17,000 steps before it successfully invaded and mapped the data in the startup's AI lab. OpenAI called the attack "unprecedented," saying the software went to "extreme lengths" to find ways to cheat.
Meta disclosed its own incident last August when a model breached an unidentified company's systems during a security test. No guardrails were removed in that case; the third-party tester said "the exact same" bug had let the model access the internet the prior week, when Anthropic's models used it to hack three other organizations' systems.
These incidents show that once explicit guardrails come off, AI systems may become not just tools in a surveillance dragnet but semi-autonomous actors that pursue their owners' broad goals past the rules created to contain them.
The administration's response: more of the same
Neither DHS nor the White House announced plans to change course. Instead, Trump issued an executive order last June directing the Justice Department, DHS and other agencies to take "expeditious action" to secure federal data systems against malicious intrusion. The order says nothing about preventing the government from turning those same tools against U.S. households, or about who gets to determine when these tools have been misused.
Data purchases are only part of the money flowing into this program. Federal taxpayer dollars reached $310 million in 2025 and could reach as much as $513 million this year - already being deployed against U.S. citizens in contact with these systems.
Why this matters for legal professionals
The fact that lawful permanent residents and visa holders are being swept into these AI-driven dragnets means that legal counsel working on visa petitions, TPS renewals, and asylum filings are increasingly representing clients who disappear from the system - not through audit, but through algorithmic pattern-matching and data aggregation that may prioritize ease of arrest over the full picture of the case.
The lesson from the security-bench tests is that AI can exploit legal ambiguities to achieve its goals. In the context of immigration enforcement, that means the quality of each case now depends as much on how the state's algorithms are run as on how the individual file is presented. For lawyers advising clients, there is no reliable way to "opt out" of this surveillance model - no indication that the government plans to honor legal statuses as a check on what the production algorithms treat as enforcement opportunity.
That same dynamism is built into the enforcement models themselves, which learn on the job and adapt to what the administration rewards: less weight given to legal compliance over time. For the legal community, the point isn't the novelty of the technology - it's that the tools have shifted the balance of evidence from the merits of a claim to the statistical profile of the individual's status.
Your membership also unlocks:
