The towers have operators. Do they have governance?

Hundreds of AI-enabled cameras now watch the southern border 24 hours a day with no human at the controls. The New York Times traced the chain: radar detects, cameras capture, a model interprets and alerts, an agent decides whether to dispatch, and the footage can travel to local, state, federal and foreign agencies. The one thing the reporting does not describe is a record of who decided what, on which clip, on what basis. Every enterprise running its own autonomous sensing will be asked for that record first.

References verified as of October 23, 2026.

What happened

On October 23, Sam Sifton's Morning newsletter at The New York Times carried new reporting by Bora Erden and Nicholas Nehamas on the surveillance towers along the southern border. The federal government has installed hundreds of new AI-enabled cameras on tall towers across the border, with plans to extend them to Florida, Puerto Rico and the northern border. The cameras see for miles in every direction and track people and vehicles around the clock. An algorithm does the watching.

The Times visited two dozen towers in California, Arizona, New Mexico and Texas and flew a drone to approximate what each tower sees. The cameras can detect people entering and leaving houses, identify and track vehicles, and follow a person or a car from one tower to the next. Some towers stand in open desert. Others stand in populated areas, where they can see into bedrooms, watch public beaches and track children on playgrounds. The Times' analysis found at least one million people in the United States and Mexico living within view of the towers, and at least four million more under an older generation of cameras the government plans to retrofit with AI.

Most of the people the Times spoke to near the towers did not know what they were. Many took them for cell towers or radar.

U.S. Customs and Border Protection told the Times that capabilities which improve an agent's awareness protect agents, migrants and the public, and that the agency uses the technology while respecting civil liberties and privacy. The agency's own privacy guidelines, as the Times describes them, say that people near the border cannot opt out of the surveillance, and acknowledge that the cameras may record people engaged in First Amendment protected activity or simply obeying the law.

The councilman's test

A Republican former city councilman in Texas gave the Times the test that matters. "We don't want to be surveilled by our government," he said, and he urged his own party to imagine the towers under the control of the other one.

Run that test on any autonomous system. Keep the system, change the operator. If the system's acceptability changes with the operator, the constraints were in the operator. Governance is the set of constraints that survive a change of operator: who may see the feed, what the model may flag, what a person may do with a flag, where the output may go, and what record all of that leaves. A system with those constraints written down and enforced can be run by either party. A system without them runs on trust in whoever holds the controls that year.

The chain, step by step

The Times' diagram shows three steps. Radar detects movement. The cameras swivel and capture the target. The AI interprets the sensor data and alerts Border Patrol. The reporting adds two more. An agent views the footage and decides whether to dispatch agents to intercept. Agents can share the footage with local, state and federal law enforcement, and with foreign governments.

Five steps. One human decision, at step four. Steps one through three run without a person, and that is the design. A tower that watches for 24 hours cannot have a person at the controls, so the person sits at the decision. That is the right place for the person. The question is what bounds the decision and what it leaves behind.

Step five is where the question lives. Sharing is a decision too, with its own authority. The reporting describes it as something agents can do, and nothing further about who authorizes a transfer, to which agency, under what rule. A clip that leaves CBP for a foreign government has crossed a boundary that a clip reviewed on a Border Patrol screen never did, and the reporting gives no sign the two are treated differently.

The record

Nowhere in the reporting is a record described. Which clips the model flagged. Which flags an agent acted on. Who decided to dispatch, and on what basis. Who received the footage afterward. The reporting may simply not have reached that record. For a government system, the record, if it exists, surfaces through oversight: inspectors general, records requests, litigation, Congress. It surfaces late and after the fact, and the privacy guidelines the Times cites cover what the cameras may see, with nothing in the reporting on what a decision leaves behind once a person acts on a flag.

For an enterprise, the record is the control. It is what the auditor reads, what the examiner asks for, and what a court subpoenas. An enterprise that runs the first four steps and cannot produce the fifth has built a sensing system and left the governance for later.

The enterprise version

Enterprises are building the same chain on their own estates, and most of them would not call it surveillance. A fraud model flags a wire and an analyst decides whether to release it. An endpoint agent quarantines a laptop and a security engineer decides whether to wipe it. A data-loss model stops an outbound email and a compliance officer decides whether it goes. An AI agent now sits in front of other AI agents, watching their sessions, and a person decides whether a session continues. The output of each decision moves: to legal, to human resources, to the regulator, to law enforcement, to the customer.

Same five steps. Same single human decision. Same sharing step at the end. And the same question arrives, from the examiner rather than the councilman. On October 13, Reuters reported that the Federal Reserve, the OCC and the FDIC are asking banks whether agent decisions can be tracked and audited, what stops an agent exceeding its authority, whether humans remain involved in key decisions, and whether an agent can be shut down. Those are four questions about a chain and its record. A bank that cannot answer them for its fraud model will not answer them for its agents, and the examiner will ask about both.

The councilman's test applies here as well. Change the operator: a new chief information security officer, an acquirer, an outsourced security operations center, one vendor's model swapped for another's. If the constraints travel with the system, the enterprise has governance. If they leave with the person who set them up, it has operators.

Owned intelligence, governed from the first call

Enterprises will rent some intelligence and own more of it. The policy is written once for the whole estate, it is enforced at the control point inside the enterprise's walls, and the evidence stays with the enterprise.

The chain an enterprise needs for a model that watches is the chain it needs for the agents it runs. A verified person, the authority that person delegated, the agent that used it, the policy that applied, the decision and the record of all of it. That chain is the product. The accountable human decides.

Related: The agents banks can block, and the ones they cannot, The CISO's guide to runtime governance for AI agents and What is a hybrid AI estate?.

Sources

• The New York Times, The Morning, "You Are Being Watched," Sam Sifton, October 23, 2026, on reporting by Bora Erden and Nicholas Nehamas. Newsletter read October 23, 2026. Article URL: paste the link Craig is posting.
• Reuters exclusive, "Banking watchdogs ramp up AI scrutiny as stability fears mount, sources say," October 13, 2026 (syndicated copy): https://www.cp24.com/news/world/2026/10/13/banking-watchdogs-ramp-up-ai-scrutiny-as-stability-fears-mount-sources-say-reuters-exclusive/. Opened October 23, 2026, for the banks post.