EU AI Act for security governance

What it is#

The AI Act is Regulation (EU) 2024/1689, laying down harmonised rules on artificial intelligence, published in the Official Journal of 12 July 2024. Regulation (EU) 2026/1744, the Digital Omnibus on AI, amended it; every citation below is to the consolidated text of 27 July 2026 ¹.

It is drafted like product-safety law. It reaches providers placing systems or models on the Union market, and deployers established in the Union. It also reaches third-country operators whose output is used there, plus importers, distributors, product manufacturers and representatives ².

Obligations scale with the tier: prohibited practices, Chapter III for high-risk systems, transparency duties, and a chapter for general-purpose AI models.

Who is in scope (decision test)#

Six questions in order. The written answer is the record.

1. AI system, or general-purpose AI model? A system infers from input how to generate predictions, content, recommendations or decisions; a model shows generality across tasks ³.
2. Which role? Provider, deployer, importer, distributor and authorised representative are defined at Article 3; product manufacturers enter through Article 2(1); one system can carry two roles ⁴.
3. Is the practice prohibited? The Omnibus added two points to Article 5(1): (ba), realistic intimate or sexually explicit material of an identifiable person made without explicit consent; (bb), material under Article 2, points (c) and (e), of Directive 2011/93/EU. Placing on the market is caught where that generation is the intended purpose, or foreseeable without safeguards. Use is caught only where the deployer uses the system for that purpose ⁵.
4. Is it high-risk? Article 6(1) catches a system that is, or is a safety component of, an Annex I product needing third-party conformity assessment. Article 6(2) catches eight Annex III areas: biometrics, critical infrastructure, education, employment, essential services, law enforcement, migration, and justice. Article 6(3) carves out a system posing no significant risk, though profiling stays high-risk and the carve-out is documented ⁶.
5. Does a transparency duty apply? Article 50 covers direct interaction, synthetic content marking, emotion recognition, biometric categorisation and deep fakes ⁷.
6. Does a model carry systemic risk? High-impact capabilities are presumed above 10^25 floating point operations of training compute; the provider notifies the Commission ⁸.

Outcome	What it means	Next step
Out of scope	No system, model, or an exclusion	Record the reason
Light duties	Article 4, plus Article 50 if applicable	Assign disclosure owners
High-risk or systemic-risk model	Chapter III, or Article 55	Start now

Obligations by article#

Security-relevant paragraphs only.

Article	Obligation	In practice	Evidence
Art. 4	AI literacy for staff and others using systems on their behalf 9	Sufficient level, none defined	Training record
Art. 9	A risk management system, iterative across the lifecycle 10	Fed by post-market data	Dated records
Art. 10, 11	Data governance on data sets, and documentation to Annex IV 11	Provenance is a control	Versioned records
Art. 12, 13	Event recording, and instructions stating accuracy and cybersecurity 12	Deployers depend on both	Logging design
Art. 14	Oversight designed in and effective in use 13	Automation bias is named	Oversight design
Art. 15	Accuracy, robustness and cybersecurity; resilience against unauthorised alteration 14	Named threats: "data poisoning", "model poisoning", "adversarial examples or model evasion", "confidentiality attacks", "model flaws"	Threat model
Art. 16, 17	Twelve lettered provider duties, and a quality management system 15	One owner per letter	Register; manual
Art. 25	Re-branding or modification makes that party provider 16	Substantial modification triggers it	Role decision
Art. 26	Instructions followed; competent overseers; input-data control; monitoring; six-month logs; worker notice 17	Suspend on Article 79(1) risk	Overseers; retention
Art. 27	Impact assessment by public bodies, public-service providers, Annex III 5(b), 5(c) deployers 18	Annex III point 2 excepted	Assessment
Art. 50	Disclose interaction; mark synthetic content and deep fakes 19	Due at first exposure	Marking build
Art. 53, 55	Model documentation and copyright policy; for systemic risk, adversarial testing, incident tracking and adequate cybersecurity protection 20	Red-teaming, by another name	Evaluations
Art. 72, 73	Post-market monitoring on an Annex IV plan; incident reporting 21	Three clocks, below	Timestamped filings

The outer limit is 15 days after the provider or deployer becomes aware. A widespread infringement, or a serious incident under Article 3, point (49)(b), is reported within two days; where a person has died, within 10 days ²².

Dates (verified)#

Date	What happens	Source
2024-08-01	Entry into force, the twentieth day after publication	23
2025-02-02	Chapter I and Chapter II, less the new Article 5 points	24
2025-08-02	Chapter III Section 4, Chapters V, VII, XII, Article 78; not Article 101	25
2026-07-27	Articles 102 to 110; the Omnibus enters into force three days after publication	26; 27
2026-08-02	General application	28
2026-12-02	Article 5(1) points (ba), (bb), and 5(1a), 5(1b); Article 50(2) for earlier generative systems	29
2027-12-02	Chapter III Sections 1 to 3 for Article 6(2) systems, less Article 6(5)	30
2028-08-02	The same for Article 6(1) systems	31

The two tracks on one estate#

The deferral split one timetable in two. One estate can carry an Article 6(2) deadline of 2 December 2027 for credit scoring, and an Article 6(1) deadline of 2 August 2028 for a safety component in a regulated product. Two dates, one inventory.

Little of the security work waits. The prohibitions already apply, and the two added points arrive on 2 December 2026, ahead of every high-risk date. Relief for systems already on the market is narrow: one is caught once its design changes significantly, and public-authority systems must comply by 2 August 2030 ³².

The work to start has no deadline of its own: what is in use, who owns it, and what proves it. Six readiness items make that concrete. The inventory is done when every use case has an ID and a named owner. The role is done when provider or deployer is written down with its reason. The screen is done when every generative use has been checked against points (ba) and (bb). The threat baseline is done when the five Article 15(5) threats map to controls. Logging is done when six-month retention is set and tested. The incident route is done when one exercise has timed all three clocks.

Mapping to NIST AI RMF and ISO/IEC 42001#

Names as published ³³ and logged ³⁴.

AI Act	NIST AI RMF 1.0	ISO/IEC 42001:2023	Gap
Art. 4	GOVERN 2 accountability structures	5.2 AI policy	Not per system
Art. 6	MAP 2 categorization performed	4.4 AI management system	No legal tier or dates
Art. 9	MEASURE 1 methods and metrics; MANAGE 1 risks prioritized	8.2 AI risk assessment, 8.3 AI risk treatment	No lifecycle iteration
Art. 10 to 12	MAP 3 capabilities; MAP 4 components	7.5 Documented information	Annex IV list; no retention floor
Art. 14, 26	MAP 3.5 oversight processes; GOVERN 3.2 human-AI configurations	5.3 Roles, responsibilities and authorities	Competence as a test
Art. 15(5)	MEASURE 2.7 security and resilience evaluated	8.2 AI risk assessment	The five named threats
Art. 27	MAP 5 impacts characterized	8.4 AI system impact assessment	Who must do it
Art. 72, 73	MEASURE 3 risk tracking; MANAGE 4 response plans	9.1 Monitoring, measurement, analysis and evaluation	The statutory clocks

Next 90 days#

Week	Action	Owner	Output
1-2	Inventory use cases, then decide provider or deployer per row	AI governance owner	Triage sheet; role decisions
2	Screen generative uses against Article 5 (ba), (bb)	Security officer	Pass or stop, dated
3-4	Classify against Article 6(1), 6(2), the carve-out	AI governance owner	Class per row, cited
4-6	Build the Article 15(5) threat baseline	Security architect role	Threat-to-control map
5-7	Set logging and retention per Article 12 and 26(6)	Platform engineering	Retention evidenced
6-8	Rehearse the Article 73 route on all clocks	Incident manager	Exercise report, timed
8-12	Confirm Article 50 marking, then plan both tracks	AI governance owner	Two-track plan

References#

1. European Parliament and Council. Regulation (EU) 2024/1689 (AI Act), consolidated text of 27 July 2026. CELEX 02024R1689-20260727. OJ L, 2024/1689, 12.7.2024. Read at the Cellar, 2026-09-05. https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:02024R1689-20260727
2. European Parliament and Council. Regulation (EU) 2026/1744 (Digital Omnibus on AI). OJ L, 2026/1744, 24.7.2026. Read at the Cellar, 2026-09-05. https://eur-lex.europa.eu/eli/reg/2026/1744/oj/eng
3. NIST. AI Risk Management Framework (AI RMF 1.0). NIST AI 100-1. https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.100-1.pdf
4. ISO/IEC 42001:2023. https://www.iso.org/standard/42001

Standards and certification names are the property of their respective owners. GRCIDE is an independent publication, not affiliated with, authorized, sponsored or endorsed by ISO, IEC, NIST or any other standards body.