Sensors, Not Servers: The Real Attack Surface of Modern Vehicles

A driveway in suburban England, before dawn. Two men, thirty meters apart, each holding a small device. Within sixty seconds, the doors unlock, the engine starts, and the car is gone — no alarm, no broken glass. This is a relay attack, and according to the Office for National Statistics, it's now behind over half of all UK car thefts. It has nothing to do with hacking software; it's manipulation of physics, radio signal capture and replay.

That's the lens this piece takes. Most automotive cybersecurity coverage gravitates toward cloud breaches and ransomware, the same language used for hacking a bank or an online store. But a modern vehicle is closer to an operational technology platform than an IT system. It's a set of physical sensors feeding physical actuators in real time, and the real question isn't "can someone breach the network****", it's "can someone make the car believe a lie about the world". That distinction matters, because the defenses look different, and so do the consequences when they fail. Understanding why requires a closer look at the vehicle's internal architecture, beginning with how its components communicate with one another.

The Nervous system: CAN bus, ECUs, and the OBD port

CAN bus (Controller Area Network) is essentially the nervous system of a modern vehicle, a two-wire network that lets dozens, sometimes hundreds, of small computers inside the car talk to each other. It was designed in the 1980s for reliability and low cost, not for security. Messages on the CAN bus carry no sender identity and no authentication, so any device connected to it is implicitly trusted by every other device on the network. That design assumed physical isolation would always keep outsiders away from the bus, an assumption connectivity has since broken down.

ECUs (Electronic Control Units) are the small embedded computers wired into that bus, each handling a specific physical function. The Engine Control Module manages fuel and ignition timing, the Break Control Module decides whether to release braking pressure, the Tire-Pressure Monitoring system reads sensor data from inside the tires. Each ECU takes a sensor reading and converts it into a physical action.

OBD-II port is a standardised diagnostic connector, required in nearly every car sold since the mid-1990s, that provides direct physical access to the CAN bus. It exists for mechanics and emissions testing, but it also functions as an open doorway: anything plugged into it communicates directly with the vehicle's internal network, with no authentication required.

This covers how the vehicle communicates internally. A modern vehicle, however, must also perceive the world outside it, which involves an entirely different set of components.

The Senses: LiDAR, radar, GPS, and cameras

These are the sensors that allow a modern, semi-autonomous vehicle to understand its surroundings, the input that feed advanced driver-assistance systems (ADAS):

LiDAR (Light Detection and Ranging) emits pulses of laser light and measures how long they take to return, building a three-dimensional map of the vehicle's surroundings. It is used for obstacle detection and emergency braking.

Radar performs a similar function using radio waves instead of light. It is effective at measuring the distance and speed of other vehicles, and is used in adaptive cruise control and blind-spot monitoring.

GPS receivers calculate the vehicle's position using things signals broadcast from satellites, essential for navigation, and increasingly, for the routing decisions autonomous systems make independently.

Cameras feed computer vision systems that recognise lane markings, traffic signs, pedestrians, and other vehicles.

CAN, ECUs, OBD, LiDAR, radar, GPS, and cameras are all physical layer components. None of them are "hacked" in the conventional sense of a software bug being exploited. Instead, they are deceived: fed a counterfeit version of reality that a downstream physical system then acts on as though it were true.

How a Vehicle is made to believe a lie

Every physical layer component described above shares one trait: it was built to trust whatever signal reaches it, not to verify where that signal came from. That gap is what makes each of the following attacks possible, and what separates them from a conventional software exploit. None of them involve breaking into a system from the outside. They involve standing close enough, in person or over radio frequency, to feed the system a convincing fake.

Entry through the diagnostic port

Through the OBD port, attackers with physical access can interpret and replay CAN bus messages, since the protocol has no authentication to prevent it. Researchers demonstrated this over a decade ago with a tool called CARSHARK, gaining real-time control over warning lights, the horn, and other vehicle functions simply by communicating with the bus directly.

Fabricating an obstacle that does not exist

LiDAR can be spoofed in two ways: relaying captured legitimate signals so that an object appears closer or farther than it actually is, or injecting entirely counterfeit signals representing an object that is not there at all. Researchers carried out the first method with roughly $50 in commodity hardware and successfully caused a vehicle's ECU to trigger emergency braking in response to a phantom obstacle. Another study targeting Baidu's Apollo autonomous driving platform used an optimization based approach and achieved a 75% success rate in forcing emergency braking, reducing a vehicle's speed from 43 km/h to a complete stop in roughly one second.

Relocating an object by 100 meters with a radio signal

Radar can be spoofed using a technique called Digital Radio Frequency Memory, which captures, modifies, and rebroadcasts radar signals to distort a sensor's reading of distance. One demonstrated experiment caused an object 121 meters away to register as just 15 meters away, not a data breach, but a fabricated reading fed directly into a vehicle's collision-avoidance logic.

Redirecting a driver who never notices

GPS is spoofed by broadcasting counterfeit but realistic satellite signals that gradually overpower the legitimate ones, since GPS receivers default to whichever signal is strongest. Researchers built a $223 spoofing device and used it to redirect 38 to 40 real-world test drivers to to an incorrect destination without anyone noticing, a 95% success rate. The same basic principle, overpowering a legitimate satellite signal with a stronger counterfeit one, had already been demonstrated five years earlier, in 2013, when researchers spoofed a luxury super-yacht off its true course, a reminder that GPS spoofing is not an automotive-specific weakness, but a structural one shared across every domain that depends on satellite positioning, maritime, aviation, and ground transport alike.

Blinding the eyes, or simply confusing them

Cameras can be blinded outright with a brief burst of laser or LED light, or deceived more subtly through adversarial physical perturbation: small, deliberately patterned stickers placed on real-world objects that cause a vision model to misclassify them. In one well-documented experiment, stickers placed on a stop sign caused it to be misread 84.8% of the time by a vision system on a moving vehicle.

The common thread

What connects all of these methods is that none require breaching a corporate network, stealing credentials, or locating a software vulnerability in the conventional sense. They require only physical or radio-frequency proximity and an understanding of how a sensor converts a physical signal into data. This represents a fundamentally different threat model from the one most cybersecurity coverage assumes, and it is the one with the most direct path to physical, safety-critical consequences.

What this actually costs

It is tempting to treat physical-layer attacks as a curiosity, interesting demonstrations confined to research labs. The financial data suggests otherwise.

A driveway crime wave, by the numbers

In the UK, relay attacks against keyless entry systems are no longer a marginal crime trend; they are now the dominant method of vehicle theft. Office for National Statistics data for the year ending March 2024 found that 58% of car thefts involved criminals replicating or manipulating a vehicle's key fob signal, up from just 14% in 2019, a more than fourfold increase over five years as keyless entry has become standard equipment. Industry data from the Association of British Insurers illustrates the financial consequences: Insurers paid out a record £669 million in motor theft claims in 2023 alone, with later reporting putting the 2024 figure at roughly £1.24 billion as both theft rates and the value of targeted vehicles continue to rise. Roughly 130,000 vehicles were stolen in England and Wales in 2024, and only about 13% are ever recovered.

Manufacturers have begun to respond. BMW, Mercedes-Benz, and Audi have introduced motion-sensing "sleeping" key fobs that stop transmitting when stationary, and Jaguar Land Rover has adopted ultra-wideband technology specifically to make relay attacks more difficult to execute. Yet a UK Home Office analysis estimates that over a quarter of vehicles on the road still carry vulnerable keyless systems, meaning the fix is arriving considerably more slowly than the exploitation.

When the cargo is not where the dashboard says it is

The freight and logistics sector tells a parallel story, with GPS spoofing in place of keyless relay. Cargo theft cost the North American industry an estimated $6.6 billion in 2025, and a growing share of that loss is now attributed not the break-ins or hijackings but to what the industry calls strategic theft: the manipulation of location data so that a truck appears to be following its assigned route on a dispatcher's screen while the actual cargo has already been diverted elsewhere. One widely cited case, referred in industry reporting as the "Santo Tequila" theft, combined GPS spoofing with fraudulent carrier credentials. Tracking systems showed two trucks moving normally right up until their delivery windows expired, by which point recovery options were essentially exhausted.

This illustrates a broader principle — a sensor reading is only useful if it can b trusted to reflect physical reality, and once an attacker controls what a sensor reports, every downstream decision built on that data becomes unreliable.

A widening gap between what's possible and what's protected

What connects the relay-attack theft wave, the GPS-spoofed cargo losses, and the laboratory demonstrations against LiDAR, radar, and cameras is one shared weaknesses: none of these systems were built to verify where a signal actually came from. A CAN bus message does not prove who sent it. A GPS signal does not prove it came from a real satellite. A LiDAR pulse does not prove it bounced off a real object. For decades, that didn't matter, because physical isolation did the verifying instead. Connectivity has quietly dismantled that assumption, and criminals haven't waited for defenses to catch up.

It is a gap the OT and industrial control systems have lived inside for years: slow, safety-constrained defenses on one side, an adversary who needs no zero-day on the other, just an understanding of how a sensor works. The car is simply the latest physical system to learn that connecting to the digital world means inheriting its risks, whether or not it inherits the defenses to match.

This article is based on original research examining security attacks and defense techniques for connected and autonomous vehicles. The full paper is available here