Zenpanik
The Red Queen's race: how Apple, Google and evolution teach systems to stop attacks earlyBy Pavel Nikolov, November 9, 2026

Part 5: The physical door

Part 5 of 6 in a deep dive on machine learning that predicts attacks and stops them early, with Apple as the main case. Start with Part 1 if you are new to the series.

Part 4 ended with a jammed lock and a question: what happens at the oldest door? Doors, vaults, safe deposit boxes and the trucks that carry gold, cash and servers were once guarded by steel and guards. In 2026 they are guarded by cameras, radar, Wi-Fi and badge readers, each running a model, and they are attacked with the same AI tools as the digital doors. This part follows the attack from the device and the exploit to the lock and finally to the physical door, where AI sits on both sides.

What is at stake in 2026

Physical assets are worth more and are moving more. Gold reached a record of $5,595 an ounce earlier in 2026, up 104% on twelve months before, and wealthy buyers want it as physical bars outside the banking system. In Switzerland, vault space is tightening because disused military bunkers that stored precious metals are being taken back into service (Financial Times via Chip). One London vault firm reported a 120% jump in demand and is adding hundreds of safe deposit boxes (Spear's).

Theft is shifting from force to deception. Verisk CargoNet counted 2,072 supply chain crime incidents in the US and Canada through September 2026, down 25%, yet losses rose 39% to $718.9 million and the average stolen load rose about 72% to $448,456 (FleetOwner, The Loadstar). Fictitious pickups, where criminals pose as a real carrier to collect a load, made up about 25% of third-quarter incidents, up from 18% a year earlier (FleetOwner). Criminals compromise business email, internet phone systems and the apps carriers use to verify identity (Insurance Journal), and a single trailer of AI chips and servers can be worth up to $15 million (Insurance Business).

Reported incidents20252,75520262,072Estimated losses, $ million2025about 5172026719Average stolen load, $2025261,1152026448,456January to September. The 2025 loss figure is derived from the reported 39% rise.
Figure 15. Fewer thefts, bigger losses. The thieves are choosing high-value loads and using deception to collect them. Source: Verisk CargoNet via FleetOwner and The Loadstar.

The Louvre showed what happens when the basics fail. A French Senate inquiry into the October 2025 theft of crown jewels worth an estimated €88 million found that only one of two cameras covering the break-in point was working, that staff lacked screens to watch footage live, and that the balcony the thieves used had been flagged as a weak point years earlier (Senate inquiry via Youboox). At the time, the museum had spent only €3 million of the €83 million set aside for its security plan (AFP).

AI on both sides of the door

For attackers, AI makes the parts that used to need skill cheap. Google saw state groups use models to map organization charts and write social engineering material (Cyber Magazine), the same research a thief needs to find who can release a load or open a vault. Forged documents and cloned voices attack the identity check in front of the lock rather than the lock itself, which is why fictitious pickups held steady while ordinary theft fell.

For defenders, the camera became a computer. Motorola Solutions says over 90% of its fixed cameras now ship with AI that detects objects and unusual activity (Memoori), and AI cameras made up about half of Hanwha Vision's fourth-quarter 2025 sales (Boannews). SECOM's security robot cocobo became the first robot from a security company cleared to travel on public roads in Japan (Envalith). Sensors learned to fuse. When drones closed Copenhagen and Munich airports in autumn 2025 (Sentrycs), the answer was classifiers trained on radar, radio and sound, combined so that no single sensor decides (Defence Online, a&s).

Confidence that the object is a droneRadarsize and speed42%Radio scannercontrol signal35%Camerashape at range47%Microphone arrayrotor sound30%Fused score91%alarm lineDrone flagged. No single sensor would have raised the alarm.
Figure 16. Sensor fusion. Each classifier alone stays under the alarm line; combined, they flag the drone. Illustrative scores; the multi-sensor approach is the one airports describe (a&s, C4ISRNET).

False alarms are the cost of every one of these classifiers. HiveWatch's 2026 benchmark found false alarm rates approaching 44% at large organizations (HiveWatch), and a guard who dismisses most alarms is, like the blocked sign-ins in Part 4, a source of biased labels. A random sample of dismissed alarms reviewed properly is the physical version of keeping the lock honest.

The physical fingerprint

Fusion works for people and vehicles as well as drones. A badge says who should be at the door. A camera sees how the person looks and walks. A phone near the reader gives off a radio signature. Wi-Fi itself can now sense: the IEEE approved its 802.11bf standard for WLAN sensing in May 2025 and published it in September, so access points can estimate presence, motion and range (IEEE, NIST). Door contacts and floor sensors count how many people actually went through. Combined, these signals form a single profile, a physical fingerprint, that is far harder to fake than any one of them.

Run 1: the badge and the person agreeRun 2: a borrowed badgeBadge readervalid badgevalid badgeCamera: appearance, gaitmatches enrolleddifferent gaitPhone near the readerusual deviceno device seenWi-Fi sensingone persontwo peopleDoor and floor sensorsone entryone entryphysicalfingerprint0.970.31door opensheld for asecond checkIllustrative scores. Each signal alone is easy to fake; five that disagree are hard to.
Figure 17. A physical fingerprint at one door. A borrowed badge passes the badge reader but not the other four signals. Schematic of the fusion approach, not a specific product.

The same idea protects a vault or a loading dock. A pickup request can be scored on the age of the email domain it came from, whether the phone number matches the carrier's history, the ID scanned at the gate, the truck's plate and the time of day. Any one can be forged with AI. All of them together, checked against what the carrier has done before, is what a fictitious pickup struggles to fake.

The fingerprint also creates its own risk. Research has shown that Wi-Fi signals alone can re-identify specific people (overview of WhoFi research), and a profile that follows someone through a building is surveillance if it is used without limits. The EU AI Act already bans some uses, such as real-time remote biometric identification by police in public spaces outside narrow exceptions. My rule would be Apple's pattern from Part 2: compute the fingerprint locally, only at doors where people have agreed to it, compare it with their own enrolled profile, and keep the verdict rather than the raw signals.

Who sells the locks

The companies below sell most of the cameras, locks, guards and armored trucks in the six markets this series looks at. Their 2025 results show where the money and the AI are going.

CompanyHomeLatest full-year revenueWhere AI shows up
HikvisionChinaRMB 92.5 billion, about $13 billion (results)Guanlan multimodal models; 12.7% of revenue spent on R&D
SecuritasSwedenSEK 155.1 billion (results)Technology and solutions about a third of sales (Q4 slides)
ASSA ABLOYSwedenSEK 152.4 billion (results)Digital access and locks; bought safe-lock maker Sargent & Greenleaf in 2025
SECOMJapan¥1,256.9 billion, year to March 2026 (overview)Face-recognition controllers and the cocobo robot (Quartr)
Brink'sUS$5.26 billion (results)Cash and valuables logistics, ATM and retail services
LoomisSwedenSEK 30.4 billion (results)Cash handling; bought a precious-metals storage site in Canada
dormakabaSwitzerlandCHF 2.79 billion, year to June 2026 (results)Access control; Swiss sales CHF 229.9 million, up 4.8% organically (release)
Motorola SolutionsUS$2.1 billion in video security and access control, up 10% (Memoori)Over 90% of fixed cameras ship with embedded AI
Hanwha VisionSouth KoreaKRW 1.34 trillion in security, up 10% (Hellot)AI cameras about half of Q4 sales; own Wisenet 9 AI chip

Two patterns stand out. The growth is in software and AI inside hardware, from Motorola's cameras to Hanwha's chips, while the labor-heavy guarding businesses grow slowly. And the companies that move valuables are buying storage: Loomis bought a precious-metals site, and ASSA ABLOY bought the maker of the locks on many of the world's safes.

What the six markets spend

There is no single dataset that measures physical security spending the same way in every country. The cleanest comparison comes from one research firm, Fortune Business Insights, which forecasts 2026 physical security markets of $45.38 billion in the US, $4.82 billion in Japan, $2.99 billion in the UK and $2.91 billion in France (Fortune Business Insights via GII).

2026 forecast, $ billion, one research firmUnited States45.38Japan4.82United Kingdom2.99France2.91South Korea, official survey: KRW 11.47 trillion in physical security revenue in 2024.Switzerland: about €1.5 billion for private security services in 2026 (IBISWorld).Japan, official survey: ¥3.5 trillion in guarding revenue in 2025, a much wider definition.
Figure 18. Market forecasts depend heavily on definitions; compare within a source, not across. Sources: Fortune Business Insights, Korea's Ministry of Science and ICT, IBISWorld, Japan's National Police Agency.

The official numbers tell a more useful story than the forecasts. South Korea's government survey put physical security revenue at KRW 11.47 trillion in 2024, up 7.3%, with growth in security cameras and camera components (Boannews). Japan's National Police Agency counted ¥3.5 trillion in revenue from the 5,210 security firms that answered its 2025 survey, and 596,985 guards, 21.6% of them aged 70 or over. Sites covered by remote alarm monitoring rose to 3.53 million while dedicated patrol cars fell to 85% of their 2021 level (National Police Agency). Fewer and older guards watch more sites, which is exactly the gap AI fills. Switzerland has no central statistic on security staff at all (Swiss federal report); IBISWorld estimates its private security services market at €1.5 billion in 2026 (IBISWorld).

Guardrails for the physical door

Verify every release on a second channel
Call the carrier or client back on a number already on file before a load, a box or a bar leaves. CargoNet says business email compromise is the main way into the most sophisticated cargo schemes (Insurance Business).
Fuse sensors, never trust one
One working camera out of two was enough to lose the crown jewels. Combine cameras, door contacts, radar or Wi-Fi so a single blind spot does not blind the site.
Fix the weak points you already know
The Louvre balcony was flagged years before the theft. A list of known gaps with dates beats a new AI product.
Keep the fingerprint local and consented
Score at the door, against the person's own enrolled profile, and keep the verdict rather than raw video or radio data.
Review a sample of dismissed alarms
With false alarm rates near 44%, the guard's dismissals are the training data. Check a random share of them.
Stage camera firmware and model updates
Roll new models to a few cameras first, as Part 4 describes for software locks.

From the door to the playbook

The physical door ends the walk inward. From the chip in a phone to the loading dock, the same pattern held: AI makes attacks cheaper and faster, and the defenses that work combine many signals and keep checking their own labels. Part 6 turns that into a playbook, with an evolutionary approach to building the models, the six findings from 2026 that held up across sources, and where I would start.

Continue to Part 6: The playbook