Two documents, two different jobs
Regulation (EU) 2024/1689, the EU AI Act, is binding law. Article 15 is titled "Accuracy, robustness and cybersecurity" and applies to high-risk AI systems. It states what a system must achieve and leaves the choice of technical measures to the provider. Provisions of the Act apply progressively, and the current text is the consolidated version on EUR-Lex dated 27 July 2026, which incorporates successive amendments. Cite that version, not the 2024 original, and recheck it before relying on any date or paragraph reference.
ETSI EN 304 223 is a different kind of instrument. It is a technical standard that sets out a detailed baseline of AI security requirements. It is prescriptive where Article 15 is outcome-based, which is why it is attractive as a source of concrete controls. It is also why the comparison needs care. A standard describes how an organisation can build and operate controls. It does not by itself define what the law requires.
That difference matters for status. The European Commission states that harmonised standards will offer legal certainty under the AI Act, and its standardisation page describes one harmonised standard aimed at Article 17 lifecycle governance requirements. Nothing in that description connects a standard to Article 15 or to ETSI EN 304 223. As of October 2026, treat any claim that ETSI EN 304 223 confers a presumption of conformity for Article 15 as unverified until it is confirmed from the Official Journal or ETSI.
What actually creates a presumption under Article 15
The AI Act text contains two routes to presumed or deemed compliance with the Article 15 cybersecurity requirements, and neither names a technical standard.
The first is certification. High-risk AI systems that are certified, or covered by a statement of conformity, under a cybersecurity scheme adopted pursuant to Regulation (EU) 2019/881, with references published in the Official Journal, are presumed to comply with the cybersecurity requirements of Article 15. The presumption applies only in so far as the certificate or statement of conformity, or parts of it, covers those requirements.
The second is Regulation (EU) 2024/2847. Where a high-risk AI system falls within its scope and the conditions in Article 12(1) of that Regulation are fulfilled, the system is deemed to comply with the cybersecurity requirements in Article 15 of the AI Act.
Both routes are limited to what the certificate or declaration actually covers. The practical consequence is that a mapping document should record, per requirement, which instrument covers it and which requirements remain uncovered. A mapping to ETSI EN 304 223 is useful evidence of alignment effort. It is not a substitute for either route.
Mapping by threat class and control location
Compliance teams usually map Article 15 against the classes of attack it is concerned with: attempts to alter a system's use, outputs or performance, and attacks on data, models and inputs. Take the exact obligation wording from the consolidated text and the provision wording from the standard. The table shows where each class of control typically operates and what evidence it can produce. It is a planning aid for building your own provision-level mapping, not a statement of how any specific provision reads.
- Manipulated inputs and prompts: Location: execution path. Runtime action: allow, block or escalate for human approval on each interaction. Evidence: the policy version, the decision and the event it applied to.
- Unauthorised actions through tools: Location: execution path. Runtime action: allow, block, rate limit or escalate on each tool call. Evidence: decision record with timestamp and approver identity where a human approved.
- Leakage of confidential data in outputs: Location: execution path. Runtime action: redact or block sensitive content in an interaction. Evidence: record of what rule applied and what action was taken.
- Abusive request volume: Location: execution path. Runtime action: rate limit by caller or tool. Evidence: threshold in force and the events that triggered it.
- Training data and model poisoning: Location: pre-deployment and lifecycle processes. A runtime layer does not cover this. Evidence: data and supply chain controls, review records and test results held elsewhere.
- Security governance and accountability: Location: organisational. Ownership, risk assessment and review triggers. Evidence: documented roles, approvals and change records.
Labelling rows as direct, partial or absent
Each mapping row needs a label and a short rationale. A row is direct when a standard provision addresses the same risk as the Article 15 obligation and a control can be shown to satisfy both. It is partial when the provision addresses part of the risk, for example it covers a technical measure but not the outcome the regulation expects, or it applies to only some of your system's components. It is absent when no provision corresponds to the obligation. Absent rows are not failures of the standard. They show where the regulation's outcome-level wording requires you to define your own control.
Two habits keep the table honest. First, leave a row unmapped until you have read both the obligation and the provision. Second, record the instrument relied on for each row: your own control, the standard, a 2019/881 certificate or a declaration under Regulation (EU) 2024/2847. Where a certificate covers only part of Article 15, the uncovered requirements still need their own controls and evidence.
The tradeoff is effort against defensibility. A coarse mapping is quick but invites questions from a notified body or regulator. A provision-level mapping takes longer, but each row can be traced to a control and to evidence.
Separating runtime controls from everything else
Once rows are labelled, classify each control by where it operates. Runtime controls act on individual interactions and tool calls in the execution path. They can allow, redact, block, rate limit or escalate for human approval, and each decision can be recorded when it is made. Pre-deployment, lifecycle and organisational controls, such as data handling, model supply chain, testing and governance roles, act outside the execution path. They need their own evidence, and a runtime layer does not replace them.
This split determines the kind of evidence an auditor can expect. For a document-based control, the evidence is a policy or procedure and records that it was followed. For a runtime control, the stronger evidence is produced by enforcement itself: the policy version in force, the decision taken, the timestamp, the event it applied to and the approver identity where a person approved an action. Evidence generated at the point of enforcement does not have to be reconstructed later. That is the difference between a control that is enforced at runtime and one that is reviewed after the fact.
One way to do this is Trussed, which enforces policy in the execution path across models, agents, tools, MCP servers, AI sites and coding assistants, and produces audit-ready evidence mapped to the EU AI Act. Its scope is individual interactions and tool calls. Controls on training data, model supply chain and pre-deployment testing remain outside it, and the mapping should say so.