ETSI TS 104 223 and the EU AI Act: what the sources do and do not say

Short answer

The two documents supplied for this explainer do not list the requirements of ETSI TS 104 223. They are an ETSI page on its Securing Artificial Intelligence committee and a European Commission page on standardisation of the AI Act. Neither one sets out the clauses, controls or obligations of the baseline standard, so this guide cannot say what the standard requires of organizations that develop, deploy or operate AI systems. It reports only what the two pages state, and it marks where they stop.

What ETSI's page says about the committee

ETSI's page describes the body that works on AI security standards. It says the Technical Committee on Securing Artificial Intelligence (TC SAI) develops technical standards to combat the security challenges it describes.

The page says these standards serve as a baseline.

It lists the stakeholders affected by the committee's work as end users, manufacturers, operators and governments. Those groups are the closest the page comes to naming who the standards are for. It does not tie any specific requirement to a role such as developer, deployer or operator.

One of the committee's stated areas of work is securing AI from attack. The page says ETSI develops cybersecurity safeguards for AI models and systems, where AI is a component in the system that needs defending.

The baseline standard as ETSI's page describes it

The ETSI page lists press releases that refer to the baseline work. It does not use the number TS 104 223 anywhere.

The first is dated 23 April 2025. The listing says ETSI unveiled a technical specification that sets a benchmark for securing AI systems against rising cybersecurity threats. The listing excerpt does not give that specification's number in the text supplied.

The second is dated 14 January 2026. The listing says ETSI announced the publication of a new standard, ETSI EN 304 223, that provides baseline cybersecurity requirements for AI models and systems.

Readers should note the numbering. The question asks about ETSI TS 104 223, a Technical Specification. The page names ETSI EN 304 223 for the January 2026 announcement and gives no number for the April 2025 specification. The documents do not say whether these are the same work at different stages or separate documents. This guide does not assume either.

The page also lists a later document that builds on the baseline. ETSI TS 104 275 V1.1.1, dated 2026-09 and shown as published, is titled as an application of the Baseline Cyber Security Requirements for AI Models and Systems principles to generative AI, and it covers preventing harm from deepfakes. That title shows the baseline has a name of its own and is being applied in further ETSI work. The page does not summarize what the baseline contains.

Other ETSI AI security documents on the same page

The page lists other published ETSI documents from the same committee. They are separate from the baseline, but they show the committee's wider output.

One is a security specification for an AI computing platform, shown as ETSI TS 104 033 V1.1.1 (2026-05).

Another is a two-part series on AI incident reporting. Part 1, ETSI TS 104 158-1 V1.1.1 (2026-03), is titled as the AI Common Incident Expression (AICIE) Global Framework. Part 2 covers the AICIE Common Container.

The page gives only titles and dates for these documents. It does not describe their content, and it does not say how they relate to the baseline requirements.

What the Commission's page says about the AI Act and standards

The Commission page explains how standards fit into the AI Act. It says the Act's requirements for high-risk AI systems must be fulfilled before placement on the market, and that these systems are monitored throughout their lifecycle.

It describes the role of standards this way: they translate legal requirements into common technical language, which simplifies compliance.

The page names the bodies that develop the harmonised standards. These are CEN and CENELEC, which it calls European Standardisation Organisations. ETSI is not named on the page.

The page says the Commission has asked CEN and CENELEC to develop standards in ten key areas. The list it gives includes cybersecurity, alongside risk management, record keeping, transparency, human oversight, accuracy, robustness, quality management and conformity assessment.

How harmonised standards work under the AI Act

The Commission page describes a sequence. Once CEN and CENELEC publish a harmonised standard, the Commission assesses whether it meets the objectives and legal requirements of the AI Act. The standard is then referenced in the Official Journal of the EU.

Use of these standards is optional. The page says providers can choose any other framework to show compliance with the AI Act. It adds that companies that apply referenced harmonised standards are presumed to comply with the legal requirements.

The page gives one example of progress. It says prEN 18286, a quality management system standard, became the first harmonised standard for AI to enter public enquiry on 30 October 2025. It says the standard is designed to help providers of high-risk AI systems comply with the AI Act's Article 17 requirements.

A news item dated 19 October 2026 says the Commission published the reference to the first harmonised standard supporting the AI Act in the Official Journal.

The link between ETSI's baseline and AI Act standardisation

Neither document states that ETSI's baseline cybersecurity standard is a harmonised standard under the AI Act. Neither says it is referenced in the Official Journal, and neither says it gives a presumption of conformity. The Commission page does not mention ETSI. The ETSI page does not mention the AI Act.

The two pages do touch in two places.

First, the Commission page says its European standardisation organisations apply an "international first" approach. International standards, when available and aligned with EU requirements, can become European harmonised standards. The page does not say that any ETSI document has been or will be taken up this way.

Second, the ETSI page says one of the committee's tasks is responding to EU policy requests related to securing AI. The page does not say which requests, or whether any concern the AI Act.

The Commission has asked for harmonised standards on cybersecurity, and ETSI has published a baseline standard on cybersecurity for AI models and systems. The documents do not connect the two, and this guide does not either.

What security leaders should check

The sources leave open points that matter before anyone relies on this standard:

• The text of the baseline itself, to confirm its requirements and which of them apply to developers, deployers and operators.
• Which ETSI document carries which number: TS 104 223, EN 304 223, or the April 2025 specification.
• Whether the baseline is referenced in the Official Journal for the AI Act. The Commission page, last updated 19 October 2026, says only that the first harmonised standard supporting the Act has been referenced.

Until those are confirmed from the standard and from Commission publications, the supplied documents support only the statements above.