AI-Defined Vehicles: Practical Gains and Implications for Automotive Cybersecurity

Key points of this blog 

• The shift from SDV to AIDV gives AI a greater architectural role. SDVs organize vehicle functions around software, while the emerging AIDV concept places AI reasoning and agent coordination closer to the center of vehicle decisions. 
• AI integration offers practical gains across the vehicle life cycle. Models and agents support automated testing, faster data analysis, and easier vehicle interactions, with benefits depending on implementation and usability. 
• Automotive cybersecurity must extend to AI decisions. OEMs and suppliers must assess how those decisions are formed, what they are permitted to influence, and how they interact with existing vehicle risks. 

“Imagine a car whose features, actions and interactions aren't just programmed — they're shaped and refined by AI models, agentic agents and data-driven processes.” 

That is how Qualcomm described its vision for the AI-defined vehicle (AIDV) in a January 2026 blog. The engineering appeal is clear: AI models and agents could help vehicles interpret information, coordinate functions, and respond to individual preferences. But giving AI a greater role in vehicle decisions also raises a security question: How do automotive OEMs ensure that those decisions remain within the vehicle’s intended operating and security boundaries? 

This two-part blog examines the shift from software-defined to AI-defined vehicles and its implications for automotive cybersecurity. Part 1 explores what the industry is building, how AI is being integrated into vehicle architectures and development tools, the engineering benefits it offers, and why these changes extend the scope of automotive cybersecurity. Part 2 examines the associated security risks and how automotive security practices can address them. 

Software enables continued vehicle development 

A software-defined vehicle (SDV) organizes vehicle functions around software, allowing capabilities to be developed, updated, and reused with less dependence on individual hardware components. The SDV architecture matters for security as well as engineering: it establishes the update infrastructure, version control, and hardware abstraction layers that any subsequent AI integration will build on. 

S&P Global Mobility’s analysis of scaling SDVs identifies hardware abstraction, centralized or zonal computing, and secure over-the-air (OTA) updates as key enablers of this flexibility. The company’s SDV Readiness Level benchmark assesses vehicle software readiness on a scale from Level 0 to Level 5, reflecting progressively greater capabilities for updating and upgrading vehicle functions. 

Volvo reaches the SDV benchmark ceiling 

Volvo Cars illustrates how these capabilities come together. In March 2026, the company announced it had achieved the benchmark’s highest rating, Level 5, making it the only legacy automaker at the time to reach that level. Volvo’s in-house HuginCore system combines electrical architecture, a core computer, zone controllers, and software to support updates across vehicle functions. 

Alongside individual automakers’ platforms, the industry is developing shared, inspectable software components. At CES 2026, the Eclipse SDV collaboration expanded its memorandum of understanding to 32 companies, up from 11 in June 2025. By jointly developing reusable open-source components, the initiative aims to make core vehicle software development up to 30% faster while improving efficiency and security. 

Why this matters for cybersecurity: These developments show how vehicle capabilities can be developed, maintained, and extended through software. As AI models and agents take on a greater role in vehicle decisions, automakers must consider how these models and agents change the architecture and what evidence is needed to assess the resulting vehicle behavior.

What makes a vehicle AI-defined? 

An AI-defined vehicle is not simply an SDV with AI features added. The distinction lies in where AI reasoning sits within the vehicle’s decision architecture. 

A 2026 paper in Frontiers in Robotics and AI describes AIDVs as vehicles “in which AI-based autonomous reasoning functions as an architectural component.” The authors distinguish learning within individual SDV components from a system in which decisions and adaptation emerge through interactions among vehicles, infrastructure, and other autonomous agents. Their definition therefore extends beyond adding AI-enabled features to an existing platform. 

The AIDV concept is still emerging. The Frontiers authors acknowledge that the fully realized AIDV they describe could remain largely conceptual, particularly the prospect of vehicles evolving autonomously through interactions. Current products should therefore be assessed by their demonstrated capabilities, without assuming that an AI-enabled cockpit or driving function represents a full transition to an AIDV. 

Table 1. Key architectural and security distinctions between SDVs and emerging AIDVs 

This blog focuses on how models and agents are taking on greater responsibility for interpreting information, coordinating functions, and influencing vehicle decisions.  

How AI integration spans the vehicle stack 

AI is not entering the vehicle as a single product or from a single vendor. Every layer of the stack, from silicon to automaker platforms to supplier software, is adding a model. Models and agents are being integrated across computing hardware, automaker platforms, and supplier systems, each at a different stage of maturity and with different implications for how vehicle behavior can be assessed or secured.  

The examples below range from hardware deployed in production vehicles to platforms intended for further development across multiple OEM programs. 

Computing hardware and driving models 

At CES 2026, Qualcomm reported that Snapdragon Ride Flex was deployed in mass-produced vehicles across eight global programs. The system-on-chip (SoC) combines digital cockpit and advanced driver-assistance system (ADAS) workloads on shared hardware. This consolidation provides computing resources for AI-enabled functions across domains, making workload isolation and resource management important architectural considerations for security teams. 

Renesas’s R-Car X5H platform follows a similar approach. In December 2025, the company announced that its 3nm SoC could run ADAS, infotainment, and gateway functions concurrently, with silicon samples and evaluation boards available to selected customers and partners. Its accompanying RoX Whitebox development kit supports generative AI and large language models for cockpit interactions. 

Alongside advances in computing hardware, suppliers are developing models for driving systems. At CES 2026, NVIDIA released Alpamayo 1, a 10-billion-parameter vision-language-action model that generates trajectories alongside reasoning traces. NVIDIA describes it as a teacher model that developers can fine-tune and distill into their autonomous-driving stacks, rather than run directly inside vehicles. 

Automaker platforms 

Automakers are integrating AI services from multiple providers into unified vehicle platforms, creating layered dependencies that security assessments must account for. 

• Mercedes-Benz’s fourth-generation MBUX integrates AI services from Microsoft and Google, while MB.OS supports OTA updates that include driver-assistance systems. Mercedes-Benz’s planned Level 4-ready S-Class pairs NVIDIA’s AI-driven decision-making with a parallel classical safety stack for future robotaxi operations. A partnership with Liquid AI also targets embedded models for on-device speech and language processing in North American vehicles, complementing existing cloud-based services. 

• The Volvo EX60 brings together its HuginCore system and Google’s Gemini assistant, combining Volvo technology with computing platforms from NVIDIA and Qualcomm. Volvo describes fleet experience, including accidents and near-misses, as a source for future improvements. Its roadmap also includes enabling Gemini to use the EX60’s cameras to answer questions about the surroundings. 

• Hyundai’s Pleos Connect platform integrates Gleo AI, an assistant powered by a large language model that searches for information and controls navigation, climate, and vehicle settings through natural-language requests. Hyundai aims to equip approximately 20 million vehicles with Pleos Connect by 2030. The company positions the platform as a step toward SDVs, with AIDVs as a longer-term goal. 

Tier-1 and supplier systems 

Bosch's AI extension platform, announced for CES 2026, uses NVIDIA DRIVE AGX Orin computing hardware to expand cockpit capabilities. The CUDA-based platform allows automakers to integrate their own models and agents, supporting applications such as voice assistance and in-cabin scene understanding.  

In January 2026, Magna announced plans to offer Hyperion-compatible electronic control units (ECUs), alongside integration, validation, and vehicle-launch services for NVIDIA’s DRIVE AV software running on DRIVE AGX Thor. Its role spans hardware and software coordination, testing, and deployment support. 

Sonatus AI Director supports deploying models from different providers, with isolated execution environments and cloud-based performance monitoring. Sonatus’s announced CES 2026 demonstrations included Michelin’s tire-wear model on NXP processors and MOTER’s driving-analysis model on Renesas silicon. This illustrates how a model supply chain can span model developers, silicon providers, deployment software, and cloud services, each of which represents a potential point of compromise. 

Representative AI offerings across the vehicle stack 

The table below brings together these examples and other offerings across the vehicle stack, summarizing their AI capabilities and architectural roles. 

Table 2. Representative AI offerings and supporting platforms across the vehicle stack 

Where AI delivers practical gains in engineering and vehicle use 

AI offers practical gains in vehicle development, data analysis, and everyday use. Some examples describe available workflows; others report improvements in specific applications. 

Reducing routine engineering work 

• Vector's CANoe AI automates a sequence of testing tasks: an agent converts a requirement into CAPL test code, executes it, analyzes any failures, revises the code, and reruns the test. Engineers can use their preferred AI endpoint and review the generated tests. 

• FPT Software’s AUTOSAR engineering platform applies agents across development tasks, with seven named roles — Insight, Architect, Modeler, Configurator, Builder, Verifier, and Trace — and support for test-case generation and optimization.  

• ETAS’s collaboration with Microsoft combines cloud-based calibration and analysis tools with agentic orchestration. ETAS says the cloud-based toolchain supports earlier testing and validation, reducing reliance on extensive on-road testing. 

• ZF Annotate addresses another labor-intensive task: preparing reference data for ADAS development and validation. It uses AI to label, classify, and track objects recorded by a separate reference sensor system. ZF cites an annotation example that reduced the timeline from 12 months to two. 

Making vehicle data easier to use 

AI can also help engineers retrieve information from large volumes of test data. A Microsoft customer case study on BMW describes development vehicles generating 5 to 10 terabytes of measurement data each week. BMW’s multi-agent system allows engineers to query telemetry in natural language, reducing their dependence on IT specialists. 

The SDV architecture provides a complementary benefit: improvements can reach vehicles already in service. Volvo announced a user experience update intended for approximately 2.5 million cars across 85 countries, including eligible vehicles built as early as 2020. Volvo Trucks reported a 24% reduction in unplanned stops associated with the expansion of OTA updates. 

Improving the driving experience without adding friction 

For drivers, additional capability is valuable when it makes a task easier. J.D. Power’s 2026 U.S. Tech Experience Index, based on responses from 68,084 new-vehicle owners, found high satisfaction and relatively few problems with features such as AI-enabled climate control. Hands-free Level 2+ driver-assistance systems did not outperform hands-on Level 2 systems on overall experience, suggesting that useful automation depends on execution and ease of use. 

Conversational features require similar care. A TU Berlin simulator study found that an informal assistant style tended to increase drivers’ cognitive distraction, particularly during conversations about personal topics. NHTSA has also proposed a simulator study to examine how modern voice command interfaces affect cognitive workload and distraction. More natural interaction can therefore introduce demands on the driver’s attention that must be evaluated alongside convenience. 

The engineering case for AI is strongest where a task becomes faster, more accessible, or easier to perform. Its growing role in vehicle decisions also defines what security teams must assess: how those decisions are formed and what they are permitted to influence. 

Why automotive cybersecurity must extend to AI decisions 

AI adds dependencies to existing vehicle risks 

AI capabilities enter an environment where existing risks remain. VicOne’s 2026 Automotive Cybersecurity Report describes the “Overlap Era,” a sustained phase in which traditional vehicle platforms, software-defined systems, and AI-enabled technologies coexist. Vulnerabilities carry forward across domains and vehicle generations, while AI introduces additional dependencies on models, data, and interactions among components. 

Practical benefits carry security tradeoffs 

Faster updates require traceability between each deployed version and the validation evidence behind its release. Agentic task completion delegates decisions about which actions to take, raising questions about whose authority the agent exercises. Personalization retains information that can shape later responses, so stored context becomes part of any investigation into how the system behaved. Edge deployment supports low latency, privacy, and offline operation, but it places inference inside vehicles with constrained computing resources and intermittent connectivity. 

Function and access shape the consequences 

An assistant retrieving information, a model estimating tire wear, and a driving model influencing a trajectory have different paths to harm. The potential impact depends on each function’s role, the systems it can reach, and the controls governing its outputs. 

The assurance question therefore extends beyond identifying which software or model is running. It also asks what influences a decision, how that decision becomes an action, and whether the resulting behavior stays within the function’s intended boundaries. 

VicOne's xPhinx addresses this challenge at the point of inference. xPhinx helps protect in-vehicle AI models and agents against prompt injection, jailbreak attempts, unsafe outputs, and sensitive data leakage, inspecting and sanitizing model inputs and outputs on-device, where connectivity, latency, and computing resources are constrained. VicOne’s xAurient maps automotive threat intelligence to vehicle components, supplier vulnerabilities, and potential attack paths, helping security teams prioritize relevant risks before vehicles reach production. VicOne’s xNexus supports operational monitoring and investigation by correlating vehicle cybersecurity events with threat intelligence across the fleet. 

Part 2 examines the evidence behind these concerns: attacks on driving models, model supply-chain compromise, and manipulation of agents with access to tools and services. It distinguishes research demonstrations from production incidents and explores how existing cybersecurity frameworks and practical controls can address the additional risks. 

Frequently asked questions 

What is the difference between an SDV and an AIDV? 

A software-defined vehicle (SDV) organizes vehicle functions around software, enabling updates and upgrades throughout its life cycle. The emerging AI-defined vehicle (AIDV) concept gives AI reasoning and agent coordination a more central architectural role in vehicle decisions. SDVs already use AI, so the distinction concerns its role rather than its presence. The AIDV concept is still maturing; current products should be assessed by their demonstrated capabilities rather than by the label applied to them. 

What practical benefits does AI offer automotive OEMs? 

AI can automate testing tasks, help engineers analyze vehicle data, prepare reference data for ADAS validation, and simplify interactions with vehicle functions. Reported gains apply to specific workflows and implementations; more automation does not automatically improve the driving experience. 

Why must automotive cybersecurity account for AI decisions? 

Models and agents introduce dependencies on training data, runtime inputs, retained context, and connected tools or services. Security assessments must examine how decisions are formed and what they are permitted to influence. VicOne’s 2026 Automotive Cybersecurity Report places these concerns within the Overlap Era, where AI-enabled functions coexist with existing vehicle technologies and their risks. 

How does VicOne help OEMs protect in-vehicle AI interactions? 

VicOne xPhinx helps protect in-vehicle AI models and agents against prompt injection, jailbreak attempts, unsafe outputs, and sensitive data leakage. It inspects and sanitizes model inputs and outputs on-device, supporting protection where connectivity, latency, and computing resources are constrained. 

How does VicOne help security teams track risks across AI-enabled vehicle platforms? 

VicOne xAurient maps automotive threat intelligence to vehicle components, supplier vulnerabilities, and potential attack paths, helping teams prioritize relevant risks. VicOne xNexus supports operational monitoring and investigation by correlating vehicle cybersecurity events with threat intelligence.