An $88 Hawaiian shirt stops Berlin's camera pilot at the bounding box

An operations screen during a behaviour-scanning pilot shows detections, tracks, alerts, whatever the dashboard has been told counts as an incident. Misses are not a row in that table. Because the recognition stages run on whatever the detector hands them, a person the detector never boxes produces no downstream output at all ^[11], so the failure Weckert is demonstrating is silent by construction.

The mechanism is unglamorous. Detectors key on edges, contrasts, limb proportions and silhouettes learned from millions of training images ^[5], and the account in Dexerto has the shirt doing two things to that: saturated colour overloading the early perception layers, and overlapping shapes breaking the outline so the model never binds arms, torso and head into one figure ^[6]. Weckert's phrasing beats most vendor copy on this point, describing a pattern that "shouts over the quiet statistical whisper that says 'person'" ^[7].

Worth being clear about provenance. This all comes from Weckert by way of Dexerto, which credits Futurism for the explanation of where in the pipeline the attack lands ^[4]. The tuning was done against YOLO, a single open-source detector, in a generative loop run until the model's confidence in seeing a person collapsed ^[8]. He says he checked the result against a live camera at Kottbusser Tor in June 2026 ^[15], roughly two months ahead of the August 2026 pilot rollout he built it to answer ^[12]. There is no independent measurement of the deployed system's recall in the material.

What gets pitched is behaviour reading. What gets built is a chain, and stage one of that chain is usually a stock detector on default weights, picked because it performs well enough on people who dress like the training images. Plenty of people do not. They wear coats that swallow the silhouette, and now some of them can buy an $88 shirt ^[1] that Weckert says was designed to hold up through fabric folds, shifting light and changing angles rather than only in controlled tests ^[14].

The sorting question for your own pipeline has two axes: which stages did you tune versus inherit, and which stages, when they fail, write a line somebody can review versus nothing at all. Inherited and silent describes stage one for almost every team shipping vision today, and that quadrant is where the risk sits unpriced. The recommendation carries a bill: treat detection as a stage you own and measure adversarially, and accept that a detector loosened enough to catch outline-breaking patterns will also bind more non-person shapes into person boxes, since it is keying on the same cues ^[13], and a human pays for each of those in review time. The thing to ask a vendor is which stages have ever been measured against someone actively trying to break them. Weckert's own framing is that the garment is "not a tool against the police, it's a test image for all of us" ^[9].