IN Brief:
- TÜV SÜD assessed generic AI tool-classification guidance for selected ISO 26262 model-based engineering use cases.
- The identified workflows cover AI-assisted model development, model review and test generation with the Simulink Agentic Toolkit.
- Individual AI outputs, whole controllers and unexamined agent configurations are not certified by the guidance.
MathWorks has expanded the TÜV SÜD certification scope associated with its IEC Certification Kit to cover generic AI tool-classification guidance for selected model-based engineering workflows under ISO 26262. The 14 October announcement identifies AI-assisted model development, model review and test-case generation. The assessment does not certify individual AI outputs, agents or complete automotive controllers.
Existing TÜV SÜD certifications included with the kit cover specified MATLAB, Simulink, Stateflow and System Composer uses. The additional assessment concerns generic guidance informed by particular use cases evaluated by TÜV SÜD, not approval of an unrestricted class of AI tools. Developers must classify their intended use and demonstrate appropriate controls within their own safety process.
Automotive functional safety places considerable weight on the integrity of engineering tools because an error introduced during design or verification can propagate into embedded software. Tool classification asks what could go wrong, whether the tool can introduce or fail to detect faults, and whether the surrounding process is able to identify those failures. The answer depends on intended use, the work products being generated and the safeguards around the activity. An automated tool that edits a model carries a different risk from one that merely formats a report.
An AI agent adds a further source of variability because its proposed changes may depend on instructions, model context and the behaviour of the underlying AI service. A plausible model modification is not necessarily mathematically or functionally correct. Even when an agent executes an existing engineering command, the command’s inputs and interpretation of results require control. Generic classification guidance therefore has to be applied alongside traceability, independent review and repeatable test evidence; natural-language instructions cannot serve as the safety argument for a generated design.
MathWorks’ Simulink Agentic Toolkit connects agents to structured modelling tasks. Its published capabilities include understanding model elements, making changes, running simulations, creating tests and associating requirements with models. Expert-authored skills guide particular engineering actions. This makes an AI-assisted workflow more directly connected to executable design assets than an isolated chatbot response, but the connection does not establish that the selected requirements or tests are complete.
The company also identifies Polyspace Agentic Toolkit for software analysis and verification work, including code robustness checks, coding-standard compliance and defect detection. These analyses provide another source of engineering evidence, while any conclusions still need to be interpreted in the context of tool configuration, project assumptions and the code being examined. A static-analysis result, whether invoked by an agent or an engineer, is not an exhaustive proof of correct system behaviour.
IEC Certification Kit supplies tool-classification and qualification materials, templates, validation resources and third-party certificates for specified development tools and workflows. Although the wider kit supports several functional-safety standards, the newly assessed AI guidance is limited to the selected ISO 26262 use cases identified by MathWorks. Neither its other standards support nor its existing tool certifications extends the AI assessment to all development activities.
In a model-based development process, traceability links requirements to implementation models, generated code and tests. Automated creation can reduce clerical effort, but it may also produce a test that reproduces the same mistaken assumption as a generated design. Reviews therefore need to establish whether the test would reveal the intended failure, not simply whether it executes successfully. Configuration control matters as well: the exact tool versions, prompts, skills and model revisions used can influence the evidence produced.
MathWorks says its agentic capabilities use the Model Context Protocol to connect to AI platforms, including GitHub Copilot, Codex, Gemini CLI and Claude Code. Compatibility through this interface does not confer TÜV SÜD certification or a common safety classification on GitHub Copilot, Codex, Gemini CLI, Claude Code or every configuration of another compatible agent. Organisations adopting the tools must still account for deployment architecture, access controls, records and their chosen model’s behaviour.
The assessed material gives developers a route to classify selected agent-assisted modelling tasks under ISO 26262. It does not establish a blanket tool-confidence level for arbitrary agents, certify a complete vehicle controller or approve the correctness of generated code and models. Any qualification claim still depends on the actual task, tool configuration, independent verification and project evidence.
