Who Certifies the AI Agent Writing Your Safety-Critical Code

A Certified Methodology, Not a Certified Answer

Engineering teams building embedded systems have spent the last two years quietly adopting AI coding agents, Copilot, Codex, Gemini CLI, Claude Code, while compliance functions have quietly worried about what happens when an auditor asks who verified the output. MathWorks has now put one concrete proposal on the table, and it deserves scrutiny rather than applause.

The company’s Polyspace Agentic Toolkit extends existing workflows into software analysis and verification, covering code robustness analysis, coding-standard compliance, and defect detection, under TÜV SÜD-certified guidance for use in engineering development lifecycles (AOL). MathWorks frames the certification explicitly around ISO 26262, the functional safety standard governing automotive and adjacent industrial control software (Engineering.com). Aerospace software sits under a comparable discipline, DO-178C, and no comparable certified toolkit has surfaced there yet. That asymmetry matters: a methodology certified for one safety-critical domain does not transfer automatically to another with its own assurance regime.

What TÜV SÜD has certified, per these accounts, is the methodology wrapped around the agent, not the agent’s output itself. That is a narrower claim than the headline implies. The certification covers MathWorks’ own toolkit and documented workflow, not every AI model that might plug in through MCP, and it says nothing about whether a given generated function is correct, only whether the surrounding analysis process is repeatable and traceable. It is also worth naming what this certification is not. ISO 26262 governs product safety. It does not satisfy ISO 42001’s requirements for an AI management system, and it does not resolve how the EU AI Act would classify an AI tool embedded in a safety component of regulated machinery, a classification that typically triggers its own conformity assessment. A vendor can be fully ISO 26262-certified on the functional safety side and still have no answer for AI-specific governance obligations layered on top.

That gap is why MathWorks is not the only model worth watching. FERC’s 2026 CIP audits took the opposite path: rather than certifying a vendor process in advance, auditors flagged AI access and shared-control risks after the fact, pushing NERC CIP-regulated entities toward tighter cyber hygiene and registration practices (Industrial Cyber; RTO Insider). NERC has also set reliability standards specifically for AI-driven data center load rather than certifying any single toolkit (ETDatacenters). That ex-post, audit-and-tighten posture reads closer to the NIST AI RMF’s continuous map-measure-manage cycle than to a one-time conformity certificate. MathWorks certifies the workflow before deployment. FERC and NERC audit the risk after deployment and adjust the standard. Neither resembles the EU AI Act’s conformity assessment model, and neither has converged with the other.

What This Actually Changes

For a functional safety team, MathWorks’ toolkit is a usable template, not a settled answer. It produces artifacts, conformance checks, defect analysis, traceability records, that an auditor can review against ISO 26262. It does not touch ISO 42001, the EU AI Act, or NIST AI RMF obligations that may apply to the same tool in parallel.

The decision in front of engineering and compliance leaders is not whether this toolkit is sufficient alone. It is whether their organization has a documented position across both lanes, product safety certification and AI governance framework, or whether it is betting on one covering the other. It does not.

---

Board record

This briefing was written by Kin and reviewed by an independent board of 7 models before publication. Ruling: CLEARED.

Seat	Reviewer	Finding
Chair · Editorial Judgment	Claude	cleared. The central argument—that methodology certification is narrower than output certification and that product-safety and AI-governance frameworks do not automatically satisfy each other—is coherent and w
Source & Claim Verification	Qwen · hosted fallback (qwen-2.5-72b-instruct)	cleared. All factual claims are supported by citations, but some sources are not directly linked to specific claims, which could be improved for clarity and rigor.
Regulatory & Framework Fidelity	Mistral	cleared. The briefing accurately distinguishes between domain-specific certifications (e.g., ISO 26262, DO-178C) and AI governance frameworks (e.g., ISO 42001, EU AI Act, NIST AI RMF), but omits explicit mappi
Technical Accuracy	Llama	cleared. The text correctly distinguishes between the certification of a process/methodology versus the certification of AI-generated output, and accurately maps the distinct regulatory regimes of ISO 26262, D
Bias, Balance & Hype Control	Gemini	cleared. The briefing effectively dissects vendor claims, clearly distinguishing between certified methodologies and certified outputs, and highlights the limitations of a single certification in a complex reg
Novelty & Non-Duplication	Grok	held. Core claims track straight off MathWorks/TÜV and FERC/NERC wire items; the methodology-vs-output and dual-lane framing is competent synthesis but not distinctive enough to clear non-duplication agains
Validation	DeepSeek	cleared. The central claim that the certification covers the methodology, not the AI’s output, is validated by the provided sources describing the TÜV SÜD certification for MathWorks’ guidance and workflow.

Sources cited: 19. Validation challenges: 0. Review cost: about $0.04. Learn how these briefings are written and verified.