NBC test: AI writes code letting a $100 drone stalk a person

A reporter at a major U.S. news outlet, working with a roughly $100 consumer drone, got publicly available chatbots to write the code that let the thing follow a specific human around a house using facial recognition. That is the setup of a new evaluation by NBC News reporter Jared Perlo, and the images in the piece, of a cheap quadcopter gliding through doorways and steering around lamps while it keeps a face centered in frame, are the part that will stick with policymakers.

The technical claim is narrower than the headline sounds, and worth stating carefully. According to the reporting, models from Anthropic, OpenAI and other providers generated software that could pilot the drone and run facial recognition against a target, on a machine anyone can order online. It is a demonstration, not a product, and the drone is a slow indoor hobbyist model rather than something operating outdoors at range. The interesting result is not that any of the individual pieces are new; face recognition, small drones and code-writing LLMs have all been around for a while. It is that a journalist can now snap them together with a chatbot instead of a specialist engineer.

That is the shift worth taking seriously. For a long time the implicit safety story around frontier models has leaned on the idea that turning capabilities into a working misuse tool still required real expertise. A published, reproducible walkthrough of a stalking rig built from consumer parts and default model access chips away at that assumption, and it lands in the same month regulators and lawmakers are already asking sharper questions about AI-generated code and about facial recognition on drones.

The honest caveat is that the NBC piece is one reporter's build, and the writeup as surfaced does not give us the full picture, which specific models did what, how much prompt work it took, whether guardrails ever fired, or how the labs plan to respond. Any of those details could change the severity of the story materially. What is not really in question is the direction. If it is this easy for a journalist to stand up a face-tracking drone with off-the-shelf models today, the interesting move for Anthropic, OpenAI, DJI-class drone makers and the state AGs watching them is to decide, quickly, whether 'write me code to follow this person' is a request the stack should be willing to fulfill at all.