Access Official Technical PDF

Share Summary:LinkedInEmail

Automotive cybersecurity is a safety engineering concern as well as an information technology challenge. Research published by Tencent Keen Security Lab in 2019 examined a Tesla Model S running Autopilot hardware 2.5 and software version 2018.6.1. The researchers investigated how access to the Autopilot ECU could affect steering control, and how camera-based machine-learning features could be influenced. The results are specific to the tested vehicle and software version, but they illustrate how connected control systems, vehicle networks and perception algorithms can intersect.

The research described three areas of concern: remote influence over steering after compromising the Autopilot ECU, physical-world manipulation of camera-based automatic wipers, and road markings that could mislead lane recognition in an experimental scenario. For engineers and technical decision-makers, the broader lesson is to assess the complete system—not only individual components or network boundaries.

Automotive cybersecurity and the Autopilot architecture

The Autopilot ECU, referred to in the research as APE, processes camera data and supports driver-assistance functions. The tested hardware included two Tegra processors, an Nvidia GPU and an Infineon Aurix controller known as the lizard brain, or LB. The researchers reported that cameras central to Autopilot were connected to APE, while LB supported functions including CAN message monitoring and communication between vehicle networks.

This arrangement matters because an ECU does not operate in isolation. APE communicated with LB through APE2LB_CAN, a logical CAN connection carried over UDP. LB acted as a gateway, extracting CAN identifiers and message data before forwarding standard CAN frames to vehicle networks, including chassis and powertrain buses. The steering system’s Electric Power Assisted Steering unit was among the components reachable through this architecture.

That creates a layered security problem: compromise of a computing module may have implications beyond the module itself if it can produce messages that other controllers accept. Network separation, message validation and gateway behavior therefore need to be considered together with the security of the ECU.

What the steering-control investigation demonstrated

The researchers analyzed the APE service responsible for transmitting vehicle control messages. They focused on the DasSteeringControlMessage, which was used for steering control in Adaptive Cruise Control and Automatic Parking Control modes. The message included a steering angle, a counter, a control-type field and a checksum. The system also used timestamps, counters and redundant CAN communication as safeguards.

According to the report, those protections did not prevent the researchers from altering steering messages after obtaining root access to APE. They described modifying the message-generation process so that the existing counter and timestamp could be reused, then sending steering commands through the vehicle’s communication path. A gamepad connected to a mobile device was used in the demonstration. The reported effects varied by driving mode; the authors described steering control while parked and in certain assisted-driving conditions.

The demonstrated chain depended on vulnerabilities that had previously enabled remote compromise of APE. The report states that this attack chain had been fixed after disclosure, and its appendix identifies firmware version 2018.24 as containing a fix for the cited root-access vulnerability. The findings should therefore be understood in their historical, tested context—not as a claim about current vehicles or software.

Camera-based perception expands the attack surface

The study also examined two features driven by camera images and neural networks: automatic wipers and lane recognition. In both cases, the researchers traced a process that began with camera data, passed through image preprocessing and neural-network inference, and produced outputs used by vehicle functions.

Automatic wipers and adversarial inputs

The tested automatic-wiper feature used images from a fisheye camera to estimate the likelihood of moisture on the windshield. The researchers generated digital inputs that raised the model’s rain score, causing the wipers to operate despite the absence of water. They then explored a physical-world approach: displaying patterns for the camera to observe. Their report describes using Worley-noise textures after other tested patterns did not produce a satisfactory result.

This is an example of adversarial machine learning: an input that appears ordinary to a person can produce an unintended output from a model. The practical engineering issue is not simply whether a model performs well on typical images. It is also how downstream controls behave when model output is wrong, ambiguous or influenced by unexpected visual conditions.

Lane recognition and road markings

For lane detection, the report describes two different outcomes. Attempts to make a lane disappear using subtle physical markings were difficult, and the researchers concluded that disabling lane recognition this way was not easy in their tests. In a separate experiment, however, small markings at an intersection were interpreted as a continuation of a lane. In the reported scenario, this misdirection could lead a vehicle in Autosteer mode toward an opposing lane.

The distinction is important: robustness against one failure mode does not establish resilience against all perception failures. The authors also noted that their tested lane-recognition behavior relied on computer vision results, and argued that the system should be able to identify when an apparent lane leads into an unsafe route. These findings concern the experimental conditions and software examined, rather than every operating environment.

Engineering lessons for connected control systems

Although the research focused on a passenger vehicle, several principles apply to industrial automation and other safety-related systems:

  • Assess trust across boundaries. Review how a compromised processor could communicate with gateways, buses and controllers, including paths that use different protocols.
  • Validate control messages at multiple layers. Counters, timestamps and checksums are useful mechanisms, but the report shows why their presence should not be treated as proof that a command is trustworthy.
  • Test perception under abnormal inputs. Camera-based features should be assessed for both digital and physical-world perturbations, with attention to the effects of incorrect model output.
  • Consider the safety response to uncertainty. Evaluate what the wider system does when lane or environmental interpretation is unreliable, not just whether the perception model succeeds in normal conditions.
  • Keep findings tied to versions and remediation. Record the exact hardware and software scope of security testing, and track fixes through deployment and verification.

Conclusion

The 2019 investigation illustrates how cybersecurity weaknesses can cross from an ECU into vehicle control, while perception vulnerabilities can originate in the physical environment rather than a network connection. For automotive and industrial engineering teams alike, the practical priority is system-level assurance: protect access to computing modules, scrutinize communication paths, test machine-learning inputs and define safe behavior when control or perception data cannot be trusted. The reported vulnerabilities and demonstrations apply to the specified historical Tesla configuration, but the engineering questions they raise remain relevant to connected, software-defined control systems.

Access Official Technical PDF

About The Author