Autonomous Vehicles and Cybersecurity: Technology, Vulnerabilities, and Outlook

At 8:57 p.m. on March 31, 2026, the Wuhan Traffic Police Headquarters began receiving reports from various parts of the city. Dozens of Baidu’s Apollo Go robotaxis had come to a stop in the middle of the road. Not on the side of the road, not in a parking lot: right in the middle of the evening traffic in a metropolis of nearly 14 million people. 

At least 100 vehicles were involved, according to an official statement posted on the local traffic police’s Weibo account. Passengers on board did not understand what was happening: the vehicle had come to a stop, the in-vehicle interface was unresponsive, and pressing the SOS button yielded no response. One passenger told local media that she waited 30 minutes before she was able to speak with an operator; many were trapped inside for nearly two hours.

The doors could be unlocked from the inside, as the police noted in their statement. Opening them, however, meant stepping out onto the traffic lanes of an elevated highway. Some passengers did so, waiting for help while clinging to the guardrail; for vehicles stuck in the center lanes, each one had to be towed away by emergency responders. Videos circulating on social media show lines of tow trucks backed up and lanes blocked for miles. CNBC reported collisions on the elevated highways caused by vehicles stopped in the lanes.

Why is this episode significant?

The suspension of robotaxis in Wuhan highlights a crucial issue for the industry, as it concerns the safety of autonomous vehicles when software, connectivity, and infrastructure operate within highly integrated architectures.

Autonomous driving involves technological, operational, and regulatory aspects that cannot be considered in isolation. Sensor fusion, OTA updates, V2X communications, levels of automation, legal liability, and cybersecurity are all part of the same ecosystem.

In this in-depth analysis, we examine the case of Wuhan as a starting point for exploring a broader issue: understanding how autonomous vehicles actually work, where their most significant vulnerabilities lie, and what challenges the industry must overcome to grow in a reliable manner.

The incident in Wuhan is not an isolated case

Apollo Go is the commercial robotaxi service with the highest number of kilometers driven in fully driverless mode: over 240 million kilometers in total, of which more than 140 million were driven without a supervisor on board, across 22 Chinese cities. In Wuhan—where Baidu announced in 2025 that it had achieved profitability per vehicle—approximately 1,000 active vehicles are in operation. The service has partnered with Uber and Lyft for its North American expansion and is in talks to enter the European market.

Following the“robo-taxi blackout,” Baidu did not issue any public statements in the hours after the incident. Apollo Go’s customer service department, contacted by several local media outlets, referred generally to “network issues in the driving system.” The police confirmed that the failure was systemic in nature—neither an isolated mechanical problem nor human error—and stated that the investigation is still ongoing.

This isn’t the first incident involving Chinese robotaxis: in August 2025, an Apollo Go with a passenger on board ended up in a construction site excavation in Chongqing; in May 2025, a Pony.ai vehicle caught fire on a street in Beijing. In both cases, no one was seriously injured, but the incidents had already sparked a debate on Chinese social media about the actual reliability of these systems.

The incident in Wuhan is on a different scale: it does not involve a single vehicle in trouble, but rather multiple vehicles coming to a standstill simultaneously in a major city. A similar incident had already occurred in December 2025 in San Francisco: a power outage had immobilized part of the Waymo fleet, causing traffic jams and sparking discussions about the operational resilience of autonomous systems in the event of a loss of connectivity.

The technical analysis published by CyberSecItalia in the days that followed identified three possible causes, which are not necessarily mutually exclusive: a cascading failure in the V2X protocol that interrupted communications between vehicles and infrastructure nodes; a faulty OTA update distributed to the entire fleet in the time slot preceding the outage; and the interruption of keep-alive signals to Baidu’s central servers—resulting in the simultaneous activation of “Safe State” on all vehicles. Different technical scenarios, same result: an architecture lacking sufficient local redundancy that, when it loses connection with the control center, does not degrade gradually. It crashes.

Apollo Go is not a niche experimental service: it has logged over 17 million cumulative rides, operates in more than 26 cities, and has active partnerships with Uber and Lyft in the Western market. The incident was covered by the BBC, TechCrunch, CNBC, and the Straits Times—not only because of the footage, but also because of the questions it raised about the structural reliability of these systems.

Why does this incident also involve cybersecurity?

Among the most plausible technical hypotheses are a faulty over-the-air (OTA) update rolled out to the entire fleet, an interruption in control signals to the central servers, and an edge case not handled by the onboard algorithms. All three causes share a common denominator: a centralized architecture lacking local redundancy, which is vulnerable to cascading failures. 

How does an autonomous vehicle work?

An autonomous vehicle processes real-time data to perceive its surroundings, plan a trajectory, and control the vehicle without continuous human intervention. The software architecture consists of three distinct but closely integrated modules:

• Perception: acquisition and interpretation of sensory data to identify vehicles, pedestrians, traffic signs, and obstacles
• Planning: Calculation of the trajectory and maneuvers based on the detected situation
• Control: translating decisions into physical commands — steering, acceleration, braking

Sensors: LiDAR, radar, cameras, and more

No single sensor technology is sufficient on its own. Autonomous vehicles use a technology called sensor fusion: the real-time integration of data from multiple sources to reduce errors and blind spots, which leverages

• LiDAR (Light Detection and Ranging): laser pulses that generate high-resolution 3D maps with a range of up to approximately 200 meters, making it ideal for urban environments. It is the most expensive component of the entire sensor suite.
• Radar: radio waves that can detect objects up to 10 km away even in fog, rain, and snow (conditions in which LiDAR performance deteriorates), but with lower resolution.
• Cameras: These provide high-resolution images for recognizing traffic signs, traffic lights, road markings, and pedestrians. They are cost-effective, but their effectiveness is limited in low-light conditions.
• GPS with HD maps: provides global positioning through continuous updates of your location on the high-definition map layer.
• Ultrasonic sensors: They have a close-range detection range of up to 5 meters, which is why they are mainly used for parking and slow maneuvers.

The Role of Artificial Intelligence

The Perception phase relies primarily on convolutional neural networks (CNNs): deep learning architectures designed to process images frame by frame. The principle is simple to describe but less straightforward to implement: the network takes the pixels of an image as input and returns the location and category of each detected object—car, pedestrian, bicycle, road sign, traffic light. Among the most widely used models in the automotive industry is the YOLO (You Only Look Once) family, designed to operate in real time even on hardware with limited resources, such as vehicle onboard computers.

However, detecting that an object is a pedestrian is only part of the problem. The system must also predict what that pedestrian will do in the next second: will they cross the street, stop, or change direction? Transformers —the architecture that also underpins language models like GPT—are used for this type of behavioral prediction. Transformers analyze temporal sequences of positions and movements, identify recurring patterns in trajectories, and produce probabilistic estimates of the future behavior of other road users. Recent studies confirm that they outperform previous recurrent models on metrics such as accuracy, AUC, and F1-score in benchmark datasets for autonomous driving.

The third element is synthetic data. Training these models requires millions of examples—and not all critical scenarios occur frequently enough in real-world driving. A system cannot wait years before it has seen enough instances of thick fog, sudden crossings, or damaged road signs. Generative AI solves this problem by creating artificial yet realistic scenarios: simulated environments with variable weather conditions, complex traffic, and different road geometries. Synthetic data can also be used to balance datasets—deliberately increasing the presence of rare scenarios or underrepresented user categories, such as children or the elderly, thereby reducing bias in perception models.

CNN, Transformers, and Synthetic Data: A Summary

• CNN (Convolutional Neural Networks): These networks recognize objects in real time from camera images. Models such as YOLO classify and locate pedestrians, vehicles, and traffic signs in just a few milliseconds per frame.
• Transformers: They analyze temporal sequences of movement to predict the behavior of other road users. The same architecture used in language models is applied to physical trajectories.
• Summary: Synthetic data: artificially generated scenarios used to train models to handle rare or dangerous situations without waiting for them to occur in reality. GANs (Generative Adversarial Networks) produce images and videos that are indistinguishable from real ones, which can be used as training data.

SAE J3016 automation levels: L0 to L5

The international standard for classifying levels of automation is the SAE J3016, updated in 2021 by SAE International in collaboration with ISO. It defines six progressive levels, but the difference between one level and another is not merely technical: it changes who is responsible for driving—and, in the event of an accident, who is legally liable.

L0–L2: The driver remains responsible at all times

At Levels 0 through 2, the driver is always the primary operator. At Level 1, the system assists with only one function—steering or acceleration/braking, but not both simultaneously. At Level 2, it intervenes on both functions at the same time: this is the case with Tesla Autopilot or Tesla FSD (Full Self-Driving), which, despite its marketing name, remains formally classified as Level 2. The driver must keep their hands on the steering wheel, remain alert, and is legally liable for any accident. The European regulation UN-ECE R-157 requires that hands return to the steering wheel within 15 seconds of inactivity: Tesla already meets this requirement through the interior camera that monitors driver distractions.

L3: Responsibility shifts to the manufacturer

Level 3 is the most significant turning point—not only technically, but also legally. When the system is active, the manufacturer is liable in the event of an accident, not the driver. The driver may take their attention off the road and only resume driving if the system explicitly requests it with sufficient advance warning.

To date, the only Level 3 system with internationally valid certification is the Mercedes-Benz DRIVE PILOT, available as an option on the S-Class and EQS. In Germany, it has been approved for speeds up to 95 km/h on highways; in Italy and most European countries, the system is physically installed in the vehicle but cannot yet be activated, because local regulations have not yet adopted the provisions of UN-ECE R-157. In China, the first mass-produced L3 models—Changan’s Deepal SL03 and BAIC’s Arcfox Alpha S—received their first national approvals in 2025.

The currently approved operating conditions for Level 3 are very specific: highways or roads with separate lanes, speeds of up to 95 km/h, favorable weather conditions, and no pedestrians or cyclists. Outside of these parameters, the system hands control back to the driver.

L4: Operates only in designated areas

Level 4 does not require an active driver, but operates only within a geofence —a specific geographic area for which the system has been certified. This is the level of robotaxis: Waymo in Phoenix and San Francisco, Baidu Apollo Go in Wuhan, and WeRide in the Middle East. Outside the operational area, the vehicle stops or hands over control. The L3 segment dominates the market in 2025 with a 46% share, while L4 is still limited to service fleets in select urban settings.

L5: No manufacturer has managed to achieve this yet

Level 5 entails full autonomy under any conditions—weather, terrain, or traffic—without a steering wheel or pedals. No manufacturer has achieved this yet. The most optimistic projections place it in the mid-2030s, but many researchers believe this estimate is conservative. The key challenge remains the long-tail problem: that endless tail of rare scenarios for which there is not yet enough training data.

The transition from Level 2 to Level 3 has raised regulatory and insurance issues that, in most jurisdictions, remain unresolved: who is liable for an accident caused by an automated system? In Italy, the Smart Road Decree and the 2018 Budget Law govern testing, but have not yet explicitly defined the civil liability of Level 3 vehicles in commercial use. OEMs seeking to deploy L3 systems in Europe must navigate country-by-country regulatory processes, secure ad hoc insurance coverage, and prepare for potential legal action in the event of a malfunction. This dynamic closely mirrors the liability issues surrounding corporate data management and GDPR compliance: when an automated system makes decisions, liability must be defined before the problem arises.

The most significant operational threshold lies between Level 2 and Level 3. Up to Level 2, the driver must remain alert with hands ready on the steering wheel. From Level 3 onward, the system takes control and the driver may cease active supervision, unless the system requests that the driver resume control.

The consumer market currently ranges from Level 2 to Level 2+. Level 4 systems are only operational in limited urban settings—such as robotaxis. Level 5, which involves full autonomy in any condition and setting, has not yet been achieved by any manufacturer.

In Europe, UN-ECE Regulation R-157, which came into effect on July 14, 2022, authorized Level 3 vehicles to operate on European roads for the first time. With the transition from L2 to L3, legal liability in the event of an accident has also shifted: it is no longer solely the driver’s responsibility, but also that of the system manufacturer. This has slowed commercial adoption by OEMs, who face insurance and regulatory implications that are still being defined in many jurisdictions.

From Argo to Waymo: A Century of Autonomous Driving

Experiments with autonomous driving date back to the 1920s. The most significant prototypes appeared in the 1980s and 1990s: the Mercedes VaMoRs van (1986), the Vamp and Vita-2 models, which in 1994 traveled 1,000 km on a Parisian highway at 130 km/h, and the European EUREKA Prometheus project.

In Italy, Alberto Broggi of the University of Parma drove Argo—a modified Lancia Thema—for nearly 2,000 km in autonomous mode in 1998, in what remains one of the most significant achievements of European research on autonomous driving.

The modern competitive phase began with the DARPA Grand Challenge (2004–2005) and gained momentum in 2009, when Google launched the project that would later become Waymo. In 2018, Waymo launched the first commercial robotaxi service in Phoenix without a safety driver on board. In Italy, testing on public roads is governed by Ministerial Decree No. 70/2018 (Smart Road Decree) and the 2018 Budget Law. In 2026, BMW obtained the first European type approval for a combination of Level 2 and Level 3 systems on a single vehicle with the new 7 Series.

The problem of rare scenarios: the main technical hurdle

Autonomous vehicles handle normal traffic well. The unresolved issue concerns the distribution of driving scenarios: most of the miles driven occur in common and relatively predictable situations. Then there is everything else—extreme weather conditions, unexpected roadwork, pedestrians with erratic paths, and unusual objects on the road. These scenarios are statistically rare but account for nearly all critical risks.

AI models trained on real-world data perform well on common scenarios. Rare cases appear infrequently in the datasets and are not learned sufficiently; Scale AI describes this as “the last 1% challenge”: not the car that can’t stay in its lane, but the one that can’t react to a situation it has never encountered before.

Technical solutions in the field include:

• Large-scale simulations in virtual environments and digital twins of cities
• Synthesized data generated using generative AI to simulate scenarios never encountered in real-world driving
• Reinforcement learning: the model learns through trial and error in simulated environments
• End-to-end approach (Wayve): the model learns directly from the sensory input stream, without any predefined concepts to detect—a structural response to the generalization challenges faced by traditional modular models

Cybersecurity and Autonomous Vehicles

A modern autonomous vehicle contains more than 100 ECUs (Electronic Control Units), receives software updates over the internet, communicates in real time with other vehicles and road infrastructure via the V2X protocol, uses 5G networks for cloud-based systems, and continuously transmits telemetry data.

Every communication channel is a potential attack vector—and, unlike a corporate server, a compromised system is moving at 80 mph down a highway.

ENISA and leading automotive cybersecurity research centers classify threats into four broad categories:

• attacks on V2X communications,
• interference with the OTA channel,
• exploiting ECUs via physical or wireless interfaces,
• attacks on the cloud-based fleet management infrastructure. 

The CAN bus: the vehicle's internal network

Before discussing external attacks, it is helpful to understand how a vehicle’s systems are interconnected. All ECUs communicate with each other via the CAN bus (Controller Area Network), an internal network protocol designed in the 1980s for simplicity and robustness—not for security. The CAN bus does not include authentication: any connected node can send messages to any other node, and all nodes receive all messages. If an attacker manages to inject messages onto the CAN bus, they can issue direct commands to the steering, brakes, engine, and any other physical system in the vehicle.

The Jeep Cherokee Case (2015): The First Remote Hack on the Road

In July 2015, researchers Charlie Miller and Chris Valasek demonstrated on camera for Wired how it was possible to take complete remote control of a 2014 Jeep Cherokee while journalist Andy Greenberg was driving it on the highway at 110 km/h. Without physically touching the vehicle, the researchers shut off the engine, applied the brakes, and took control of the steering—all via the Internet, exploiting a vulnerability in the Uconnect infotainment system. The point of entry was the infotainment system’s chip, whose firmware was remotely rewritten to inject arbitrary commands onto the internal CAN bus. Chrysler recalled 1.4 million vehicles. That incident marked the beginning of automotive cybersecurity as a serious discipline.

In 2018, Miller and Valasek demonstrated a second generation of CAN bus attacks that was even more sophisticated: instead of compromising a single ECU and using it to send messages, they simultaneously attacked the legitimate ECU, causing it to send conflicting commands—effectively bypassing the system’s natural defense mechanism. The result: complete control over the parking brake, steering, and cruise control at any speed.

V2X Attacks: Falsifying the Outside World

The V2X protocol enables vehicles to exchange real-time information with other vehicles, with road infrastructure (such as traffic lights and barriers), with pedestrians via smartphones, and with the traffic management network. These exchanges take place over standardized radio channels—DSRC at 5.9 GHz in Europe or C-V2X over a 5G network—and serve as the vehicle’s “extra eyes” on the surrounding traffic.

The problem is that most V2X messages in circulation today are not robustly authenticated or encrypted. A hacker, using a standard radio transmitter, can inject fake messages into the network without being detected:

• GPS Spoofing: A malicious transmitter superimposes a fake GPS signal onto the satellite signal, causing the vehicle to calculate an incorrect position. Researchers at the University of Austin demonstrated in 2013 how to hijack a luxury yacht in the Mediterranean by altering its GPS coordinates without the navigation system detecting the anomaly. The FusionRipper project , published on arXiv, subsequently demonstrated how a “stealthy” GPS spoofing attack can deceive Baidu Apollo’s multi-sensor fusion system with a success rate of over 90%.

Falsification of BSM (Basic Safety Message) messages: BSMs are packets sent every 100 ms by each vehicle to the others, containing location data.

Centralized architecture and cascading risk

When a single control node manages hundreds of vehicles simultaneously, a malfunction in that node affects the entire fleet. The case of Wuhan is the best-documented example of this. The technical hypotheses most frequently cited by experts include:

1. A faulty OTA update that triggered "Safe State" on all vehicles simultaneously
2. An interruption in the keep-alive signals to the central servers, which triggered the scheduled emergency shutdown
3. An edge case not accounted for by the algorithms, which, when propagated from the cloud, caused a decision loop in the onboard systems

In all three scenarios, the lack of local redundancy prevented the vehicles from shutting down in an orderly manner—autonomous pulling over, unlocking of doors, emergency communication—turning a technical failure into a massive urban gridlock.

Self-driving cars and road safety: what the data shows

In the U.S., as of November 2025, a total of 5,202 incidents involving autonomous vehicles had been recorded, at a rate of 9.1 incidents per million miles driven, compared with 4.1 for vehicles driven by humans. In absolute terms, that is more than double.

An analysis of liability, however, changes this interpretation. According to NHTSA data for the 2021–2025 period, autonomous vehicles were found to be at fault in only 4% of accidents involving other road users. The high absolute frequency reflects three factors: concentration in dense urban areas, mandatory reporting even for minor collisions, and the tendency for autonomous vehicles to be rear-ended—rather than to rear-end others—because their behavior is unpredictable to other road users, who are accustomed to reading the intentions of a human driver.

Algorithmic bias in perception systems: a real-world problem

Recent studies have found that pedestrian detection rates are 7–10% lower for people with dark skin and 19–26% lower for children compared to fair-skinned adults. Similar results have been documented for women and older adults.

The cause is structural: training datasets, historically built on non-representative populations, generate models that behave inconsistently toward user groups that are underrepresented in the data. In a system that makes real-time decisions on the road, this is no minor issue.

The public debate on autonomous driving is often dominated by the so-called Trolley Problem —the philosophical dilemma of who to save in the event of an unavoidable accident. Researchers at the Brookings Institution and the University of Hamburg agree that this is a misplaced problem: real-world vehicles never operate in such deterministic scenarios. The most concrete and urgent ethical risk, however, is that of algorithmic bias in training data—a solvable engineering problem, not a metaphysical dilemma.

The Robotaxi Market: Key Players, Statistics, and Projections

The robotaxi market, valued at approximately $1.95 billion in 2024, is projected to reach $43.76 billion by 2030, with a CAGR of 73.5%. Goldman Sachs estimates that the robotaxi sector alone will exceed $25 billion by the same year. BCG forecasts 120,000 units in circulation in European cities alone by 2030.

Cruise (GM) suspended its robotaxi service at the end of 2023 following a series of high-profile incidents that jeopardized its regulatory approval status. The case has become a benchmark in the industry for how the management of a reputational crisis interacts with regulatory approvals, with repercussions that extend far beyond the individual technical incident.

The V2X cybersecurity market is growing in tandem: from $1.43 billion in 2022, it is projected to reach $4.48 billion by 2029, with a CAGR of 17.7%. Spending on cybersecurity for connected vehicles is no longer an optional item in OEM budgets.

The direction for the coming years

The priorities identified by the industry focus on four key areas:

• Graceful degradation: Vehicles must be able to continue operating in a reduced capacity—safe stopping, unlockable doors, active emergency communication—even in the event of a partial failure. The “fail-safe” paradigm (complete shutdown) is gradually being replaced by the “fail-operational” paradigm (orderly degradation).
• Offline-first architecture: reduced reliance on central servers, with autonomous fallback logic built into the vehicle. The cloud optimizes, it doesn't control.
• Zero Trust for the fleet: every ECU and every vehicle is treated as a potentially compromised endpoint, with continuous authentication and strict segmentation. This is the same approach that professional IT support services use for distributed corporate networks.
• Security by Design: security built into hardware and software from the design stage, rather than added later as a patch. A mandatory requirement for EU type approval starting in 2022, in accordance with the ISO/SAE 21434 framework.

Level 5—full autonomy in any scenario—remains a distant goal. The realistic focus for the next ten years is Level 4 in an ever-wider range of areas, with Level 2 and Level 3 systems rapidly gaining traction in the consumer market.

The 5 principles of cybersecurity that autonomous vehicles are learning the hard way

• Redundancy: No critical system should depend on a single node
• Continuous authentication: every component that communicates with others must verify its own identity
• Signed updates: no software changes without cryptographic verification of the source
• Local fallback: in the event of a loss of connectivity, the system must be able to operate in degraded mode
• Behavioral monitoring: detecting anomalies in real time, before they turn into failures or incidents

These same principles apply to the management of corporate networks, cloud systems, and distributed endpoints. To schedule an assessment of your infrastructure, please contact our IT team.