About the author
Written by Alexis Hirschhorn, trainer at Abilene Academy. Certified ISO/IEC 42001 Lead Implementer, Certified Artificial Intelligence Professional, CISSP, CISM, CISA, ISO/IEC 27001 Senior Lead Auditor and PECB Certified Trainer. About 25 years of hands-on experience in information security, cloud security and AI governance, advising multinationals, government bodies and international organisations. Profile: https://www.abileneacademy.ch/en/about/team/alexis-hirschhorn
As of 18 November 2026, ISO/IEC 27091 is at stage 50.20 (FDIS ballot), with publication estimated for February 2027. It is guidance: you cannot be certified to it.
Methodology
Based on the ISO catalogue and the Draft International Standard; we will re-verify against the published text and log changes below.
1. Key takeaways
ISO/IEC 27091 is a new, non-certifiable guidance standard on privacy risk in AI systems, and it is in its final approval vote. These five points are what most teams need to know today.
- 3 February 2023: the project was registered in ISO/IEC JTC 1/SC 27. It is Edition 1, a brand-new standard that replaces nothing, so no certificates are affected and there is no transition.
- 2 November 2026: ISO/IEC FDIS 27091 entered its final eight-week ballot (stage 50.20). The vote is estimated to close around 28 December 2026.
- February 2027 (estimated): publication is most likely in February 2027, within an estimated range of mid-January to April 2027.
- 14 October 2026: the sister standard ISO/IEC 27090:2026 on AI security threats was published. It points to 27091 for the privacy side.
- As of 18 November 2026: no 27091 certification, auditor or transition deadline exists, and there is no "27091:2025". Use it inside an ISO/IEC 42001 AIMS or ISO/IEC 27701:2025 PIMS.
2. Status tracker
ISO/IEC 27091 has passed its enquiry stage (DIS) and is now being voted on as a Final Draft International Standard. The table below separates what ISO has confirmed from what we estimate.
Stage: 20.00
Stage: 40.20
Stage: 40.60
Stage: 50.00
Stage: 50.20
Stage: 50.60
Stage: 60.60
What happens next. National standards bodies are voting on the FDIS until the ballot closes, which we estimate will be around 28 December 2026. ISO/IEC 27090 took eight weeks from the close of its vote to publication, so we estimate that 27091 will be published in February 2027, with an estimated range of mid-January to April 2027. In the rare case that the draft is referred back to the committee, we estimate publication would slip to the second half of 2027. The FDIS is not on sale; buy the published standard from ISO or your national standards body.
Last verified against the ISO catalogue: 18 Nov 2026
Gauge of ISO/IEC 27091 on the ISO stage scale from 00 to 60, as of 18 November 2026. Current stage 50.20: FDIS ballot, opened 2 November 2026, day 16 of 56. Confirmed milestones: 20.00 project registered on 3 February 2023; 40.20 DIS ballot opened 3 December 2025; 40.60 DIS ballot closed 26 February 2026; 50.00 FDIS registered 13 July 2026; 50.20 FDIS ballot opened 2 November 2026. Estimated: 50.60 ballot closes around 28 December 2026; 60.60 publication in February 2027 (range mid-January to April 2027). If referred back to committee, publication would slip to the second half of 2027 (estimated).
3. What ISO/IEC 27091 is and is not
ISO/IEC 27091, Cybersecurity and Privacy - Artificial Intelligence - Privacy protection (Edition 1, JTC 1/SC 27), a guidance standard, helps organisations identify and treat the privacy risks that are specific to AI systems and machine-learning models. It is not a management-system standard, and no one can certify you, your product or your staff against it.
ISO's public abstract says the standard "provides guidance for organizations to address privacy risks in artificial intelligence (AI) systems and machine learning (ML) models" (ISO, ISO/IEC FDIS 27091 catalogue entry). The FDIS runs to 35 pages. The committee is ISO/IEC JTC 1/SC 27, and the UK mirror committee is BSI IST/33/5.
ISO/IEC 27091 is: Guidance ("should" recommendations)
ISO/IEC 27091 is: A method for analysing AI privacy risk
ISO/IEC 27091 is: A catalogue of AI-specific privacy threats and mitigations
ISO/IEC 27091 is: Usable by organisations of any size or sector that build or use AI
Datasheet for ISO/IEC 27091, Edition 1, at final draft (FDIS) stage. Title: Cybersecurity and Privacy - Artificial Intelligence - Privacy protection. Committee ISO/IEC JTC 1/SC 27, UK mirror BSI IST/33/5. 35 pages. A guidance standard, not a management-system standard, and not certifiable. It replaces nothing, so there is no transition and no certificates are affected. For organisations of any size or sector that develop or use AI. Use it to identify and treat AI privacy risks, as input to ISO/IEC 42001 or ISO/IEC 27701:2025, and as evidence for DPIAs and AI Act documentation. It is not a harmonised standard, not a replacement for a DPIA or legal advice, and not a guide to whether to deploy AI. There is no 27091 certification, transition deadline or ISO/IEC 27091:2025.
Who it is for. Organisations of any size or sector that develop, procure, fine-tune, deploy or operate AI systems processing personal data, from a start-up building a recommender to a bank buying a credit-scoring model. The main readers will be DPOs, privacy engineers, CISOs, AI governance leads, ML engineers and their consultants.
What it leaves out. The standard does not tell you whether to deploy AI at all. That business decision stays with you, as does the question of lawfulness.
“Use 27091 as a lens, not a certificate. It helps you ask the questions an auditor may not ask: Can this model be prompted into revealing training data? Can its outputs identify someone? Can a false personal claim be generated and circulated?” Alexis Hirschhorn
For a short answer to the most common search, see Is ISO/IEC 27091 certifiable?.
4. Why it exists
ISO/IEC 27091 exists because AI systems create privacy risks that classic privacy methods were not designed to find. Personal data can end up inside a model, be inferred from its outputs, or be invented by it.
Traditional privacy risk assessment, such as the approach in ISO/IEC 29100 and ISO/IEC 27557, assumes that personal data sits in places you can map, minimise, secure and delete. AI breaks that assumption in four ways (for the general picture, see what makes AI risk different from traditional IT risk).
“A traditional database is like a locked filing cabinet: you know which drawer holds which person’s file. An AI model is more like a person who has read thousands of files and can be asked questions. Even if you remove the cabinet, the model may still remember fragments, infer missing details, or combine facts in ways nobody anticipated.” Alexis Hirschhorn
Memorisation. Large models can retain fragments of training data, such as names, contact details or rare records. That data is spread across model parameters rather than held in a database, and a well-crafted prompt may bring it back out.
Inference. A model can derive facts nobody gave it, such as health conditions from purchase history, and an attacker may infer whether a person's record was in the training set.
Re-identification. Data pseudonymised or "anonymised" before training can sometimes be re-linked to individuals through outputs, embeddings or other sources. This is why European regulators no longer accept a model as anonymous on its developer's word (see section 10).
Hallucinated personal data. Generative systems can produce confident, false statements about real people, such as a fabricated conviction. This is misbehaviour rather than an attack, but it still raises issues of accuracy, harm and rectification.
5. Structure walkthrough
Based on the DIS, the core of ISO/IEC 27091 is three clauses (a privacy analysis framework, model-level privacy and lifecycle privacy) plus two annexes. This is our paraphrase of expert reviews of the DIS; buy the published standard for the authoritative wording.
Concentric diagram of ISO/IEC 27091 based on the DIS. Outer ring, clause 5: the privacy analysis cycle of identify, analyse, evaluate and treat, then repeat. Middle ring, clause 7: the AI system lifecycle of design, development and deployment. Core, clause 6: the AI model and its threats, membership inference, training-data extraction, model inversion, poisoning and insider risk, with mitigations such as access limits, pseudonymisation and input and output filtering. Annex A supports the privacy analysis; Annex B is an AI use-case template.
Clause 5: A framework for privacy analysis of AI systems
Clause 5 sets out a four-step flow: identify, analyse, evaluate and treat privacy risks. The shape will be familiar from ISO/IEC 27557. The value is in tailoring each step to AI: where personal data enters (training, fine-tuning, retrieval, prompts, logs), what the model could reveal or infer, and who could exploit it.
Clause 6: Privacy of AI models
Clause 6 treats the model itself as a source of privacy risk: membership inference, training-data extraction, model inversion, data poisoning and insider risk. Its mitigations are generic categories of control: access restrictions, anonymisation or pseudonymisation of training data, and input and output filtering.
Clause 7: Privacy across the AI system lifecycle
Clause 7 covers the wider system through design, development and deployment: purpose and data needs at design; collection, labelling, training and testing in development; interaction, logging, monitoring and updates in deployment. This is where privacy by design becomes concrete for AI teams.
Annexes A and B
Annex A supports the clause 5 analysis, and Annex B provides a template for describing an AI use case. We do not reproduce it; section 12 shows how to build a use-case profile in your own format.
Beyond attacks: misbehaviour and AI for privacy
The DIS treats misbehaviour, such as hallucinated personal data, as a privacy issue alongside attacks. It also discusses AI for privacy: consent management, privacy-enhancing technologies (PETs), federated-learning security and NLP for analysing privacy policies.
6. The AI privacy threat catalogue in plain language
Based on the DIS, ISO/IEC 27091 lists about eight AI privacy threats. Below are the six most discussed in expert reviews, each with one example, in our own words rather than the standard's.
The scenarios are a support chatbot fine-tuned on past transcripts, an HR screening model trained on historic hiring data, and a RAG assistant searching a vector store built from CRM records and emails.
Map of an AI system: training and fine-tuning data feed the AI model; a knowledge base feeds it through retrieval; users send queries and receive answers; prompts, logs and the vector store are kept. Six numbered threats are placed where they strike. 1 Membership inference on queries (HR screening example); 2 training-data extraction on answers (chatbot); 3 model inversion on queries (HR screening); 4 data poisoning on training data and the knowledge base (RAG assistant); 5 insider risk on logs and the vector store (RAG assistant); 6 hallucinated personal data on answers, a misbehaviour rather than an attack (chatbot). Each comes with a treatment direction.
Membership inference
What goes wrong: someone works out whether a specific person's data was used to train a model, usually by studying how confidently the model responds to that person's record.
Example (HR screening): a former applicant queries the model with their own profile, and the scores reveal their file is in the training data despite a promise to delete it.
Treatment direction: limit who can query the model and how often, reduce overfitting, and bound what any single record contributes.
Training-data extraction
What goes wrong: a model reproduces pieces of its training data word for word when given the right prompt.
Example (chatbot): asked for "an example of a complaint about a late refund", the bot outputs a near-verbatim past transcript with another customer's name and phone number.
Treatment direction: remove or pseudonymise direct identifiers before fine-tuning, deduplicate training data, and filter outputs for personal data patterns.
“If your chatbot was fine-tuned on real customer transcripts, assume that a determined user may try to make it repeat one. The question is not whether your model is perfect; it is whether you can show what you did to prevent verbatim leakage.” Alexis Hirschhorn
Model inversion
What goes wrong: an attacker uses a model's outputs to reconstruct sensitive attributes or representative features of the people in its training data.
Example (HR screening): probing the model with synthetic profiles, an analyst reconstructs which characteristics drive rejection, exposing proxies for health or ethnicity in the historic data.
Treatment direction: restrict access to detailed scores and confidence values, return coarse outputs where possible, and test models for attribute leakage before release.
Data poisoning
What goes wrong: someone plants manipulated data in a training set or knowledge base so that the system behaves in a way that harms individuals or leaks their data.
Example (RAG over customer data): a malicious email in the knowledge base carries hidden instructions that make the assistant include other customers' contact details in its answers.
Treatment direction: control and validate what enters training sets and knowledge bases, track where data came from, and monitor outputs for anomalies.
Insider risk
What goes wrong: people with legitimate access to models, training data, embeddings or logs misuse that access.
Example (RAG over customer data): an engineer with administrator rights searches the vector store and prompt logs for information about a well-known customer.
Treatment direction: apply least-privilege access to model artefacts and logs, separate duties, log administrator activity and set retention limits.
Hallucinated personal data
What goes wrong: the system generates false but plausible information about a real, identifiable person.
Example (chatbot): the bot tells a customer that a named employee "was dismissed for mishandling accounts", which is fabricated.
Treatment direction: ground answers in verified sources, block generation of personal claims outside defined use cases, and prepare a process for rectification and erasure requests about model outputs.
“For generative AI, accuracy is a privacy control. A system that invents personal claims needs the same discipline as a system that mishandles personal data: source grounding, output restrictions, human review, and a route for people to challenge what was said about them.” Henri Haenni
The remaining threats and the standard's own descriptions are in the standard itself. For the security-side catalogue, see our ISO/IEC 27090:2026 guide.
7. Using 27091 inside an ISO/IEC 42001 AIMS
If you run, or are building, an AI management system under ISO/IEC 42001:2023, ISO/IEC 27091 is a source of privacy-specific input for your AI risk assessment, your AI system impact assessment and your data controls. 42001 tells you that you must assess AI risks and impacts. 27091 helps you do the privacy part of that assessment properly.
ISO/IEC 42001:2023 is certifiable but does not go into depth on threats such as memorisation or membership inference. 27091 fills that gap without changing your certification scope.
Hub diagram with ISO/IEC 27091, a non-certifiable guidance standard, at the centre. Solid links: it feeds ISO/IEC 42001:2023 AIMS (certifiable; AI risk and impact assessment, data controls) and ISO/IEC 27701:2025 PIMS (certifiable; risk assessment and Statement of Applicability; 27701:2019 certificates end October 2028). Dashed link: it supports evidence under GDPR and EDPB Opinion 28/2024 (case-by-case anonymity test for models, DPIA input). Dotted link: it supports EU AI Act data-governance and risk documentation but gives no presumption of conformity. It is cross-referenced with sister standard ISO/IEC 27090:2026 on AI security (published 14 October 2026); related guidance is ISO/IEC 23894, 42005, 29100 and 27557.
ISO/IEC 42001 element: AI risk assessment (clause 6.1.2)
ISO/IEC 42001 element: AI risk treatment (clause 6.1.3)
ISO/IEC 42001 element: AI system impact assessment (clause 6.1.4; Annex A.5)
ISO/IEC 42001 element: Data for AI systems (Annex A.7)
ISO/IEC 42001 element: AI system lifecycle (Annex A.6)
Practical tips. Reference 27091 in your risk methodology as guidance, not a requirement, and keep one AI risk register with privacy threats tagged. Auditors assess you against 42001, not 27091.
“A useful test is simple: if your AI risk assessment would look identical if you replaced the AI system with a conventional database, it is probably not an AI risk assessment. 27091 pushes teams beyond that point.” Alexis Hirschhorn
For the wider implementation picture, see our ISO 42001 executive playbook, and for what auditors ask to see, the ISO 42001 audit evidence playbook. For implementation skills, see ISO 42001 Lead Implementer. For the audit perspective, see ISO 42001 Lead Auditor.
8. Using 27091 inside an ISO/IEC 27701:2025 PIMS
In a privacy information management system under ISO/IEC 27701:2025, ISO/IEC 27091 supplies the AI-specific privacy risks for your PIMS risk assessment and the reasoning behind your Statement of Applicability (SoA) decisions for AI processing. It does not add controls you must certify against.
ISO/IEC 27701:2025 is now a standalone, certifiable standard that no longer needs ISO/IEC 27001 alongside it. Once AI systems are part of your processing, generic privacy risks are not enough. 27091 helps in three places:
- Risk identification. Add AI-specific risks, such as extraction, inference and hallucinated personal data, for each AI processing activity.
- SoA justifications. Use 27091 as a recognised reference for controls such as output filtering or pseudonymising fine-tuning data.
- Controller and processor roles. Show which party (model provider, host or deployer) can treat which risk, and reflect it in contracts.
The 27701 transition. This is separate from 27091 but often confused with it. Certificates against ISO/IEC 27701:2019 end at the end of October 2028, so plan your move to the 2025 edition with your certification body. Folding 27091-informed AI risks into that work is efficient but optional. 27091 itself has no transition, because it replaces nothing.
For what changed between editions, see the main differences between ISO/IEC 27701:2019 and 2025. For the full picture of the new edition, see our ISO/IEC 27701:2025 guide.
9. ISO/IEC 27090 vs 27091
ISO/IEC 27090 deals with security threats to AI systems, and ISO/IEC 27091 with privacy risks in them. Most organisations will want both.
Aspect: Focus
Aspect: Threats
Aspect: Primary audience
Aspect: Status (18 Nov 2026)
Aspect: Certifiable?
Aspect: Relationship
Venn diagram. ISO/IEC 27090:2026, AI security: security threats to AI systems, 13 threats in the draft, for CISOs and security architects, published 14 October 2026. ISO/IEC 27091, AI privacy: privacy risks in AI systems and ML models, about 8 threats in the DIS including hallucinated personal data, for DPOs and privacy engineers, at FDIS stage with publication estimated for February 2027. Shared: both are guidance and not certifiable, both come from ISO/IEC JTC 1/SC 27, data poisoning matters to both, and they cross-reference each other.
How to use them together. Run one combined AI threat analysis. Where a threat appears in both, such as poisoning, record it once and assess both its security and privacy impact. For a deeper side-by-side, see ISO/IEC 27090 vs 27091.
10. GDPR and EDPB Opinion 28/2024
ISO/IEC 27091 is not a GDPR compliance tool in its own right, but its model-level threats are the technical side of a question European regulators now ask directly: is your AI model actually anonymous? The European Data Protection Board's Opinion 28/2024 of 17 December 2024 says that AI models trained on personal data cannot always be considered anonymous, and that anonymity must be assessed case by case.
A controller claiming anonymity must show that personal data cannot be extracted from the model, or obtained through queries, with means reasonably likely to be used. Membership inference, extraction and model inversion are exactly those routes, and a 27091-style analysis gives you a structured way to test and document them.
“Under European privacy expectations, anonymity is not a label you apply before training. It is a property you must be able to demonstrate after training, through testing, access controls, and evidence that personal data cannot reasonably be extracted or inferred.” Marc Bouvier
27091 also supports other GDPR obligations:
- Data protection impact assessments (Article 35). The threat analysis and lifecycle view can feed the risk section of a DPIA for an AI system. Our answer on running a DPIA for an AI system shows one way to structure this.
- Data protection by design and by default (Article 25). Clause 7's lifecycle approach turns this duty into concrete decisions at design, development and deployment.
- Accuracy (Article 5(1)(d)) and data subject rights. Treating hallucinated personal data as a privacy risk helps you plan how to handle rectification and erasure requests that concern model outputs.
If you run an ISO/IEC 27701 PIMS, it already structures much of this evidence; see how ISO/IEC 27701 supports GDPR compliance.
What it does not do. It does not establish a legal basis, is not a GDPR certification mechanism and does not replace the EDPB's case-by-case assessment.
11. EU AI Act timing: where 27091 helps and where it doesn't
ISO/IEC 27091 can support your EU AI Act data-governance and risk documentation, but it is not a harmonised standard and gives no presumption of conformity. Treat it as good-practice evidence, not as a compliance route.
The EU AI Act (Regulation (EU) 2024/1689, as amended by the Digital Omnibus, Regulation (EU) 2026/1744) applies in stages.
Date: 2 Dec 2026
Date: 2 Dec 2027
Date: 2 Aug 2028
Timeline drawn to scale from mid-2024 to 2028, with today, 18 November 2026, marked. Past: EDPB Opinion 28/2024 on 17 December 2024; ISO/IEC 27090 published 14 October 2026; ISO/IEC FDIS 27091 ballot opened 2 November 2026. Ahead: EU AI Act new prohibitions 2 December 2026; ISO/IEC 27091 publication estimated for February 2027; EU AI Act Annex III high-risk obligations 2 December 2027, about ten months after estimated publication; Annex I high-risk obligations 2 August 2028; ISO/IEC 27701:2019 certificates end at the end of October 2028.
Where 27091 helps.
- Data governance and risk management. Its analysis of training-data risks, provenance and pseudonymisation supports providers' documentation, and deployers can use it as a privacy input to fundamental-rights assessments.
Where it doesn't.
- It gives no presumption of conformity. Only harmonised standards cited in the Official Journal do that, such as EN 18286:2026 for AI Act quality management.
- It does not address transparency duties, human oversight, registration or general-purpose AI obligations.
“For high-risk AI systems, the standard can help you document the privacy side of risk management. But it will not classify your system, replace a conformity assessment, or answer the question regulators ultimately care about: what harm could this system cause, and what did you do about it?” Tania Postil
For every application date, see the EU AI Act application timeline. Deployers with fundamental-rights duties should also read what a fundamental rights impact assessment involves, and Swiss readers whether the EU AI Act applies to Swiss companies. Track the AI Act's moving parts on our EU AI Act hub.
12. A 90-day starter plan
You can build a working AI privacy risk process in about 90 days without waiting for publication, starting from the structure the DIS describes and aligning with the final text later. The plan assumes a mid-sized organisation with a handful of AI use cases.
A 13-week, roughly 90-day AI privacy plan in five phases, drawn to scale. Weeks 1 to 2: AI inventory of all AI systems with personal-data flags. Weeks 3 to 4: use-case privacy profile for each priority use case. Weeks 5 to 8, the longest phase: privacy risk assessment, producing risk register entries linked to the DPIA. Weeks 9 to 11: treatment plan with owners and dates. Weeks 12 to 13: monitoring plan and review calendar. Re-run when a model, dataset or use case changes.
Weeks: 1–2
Weeks: 3–4
Weeks: 5–8
Weeks: 9–11
Weeks: 12–13
Test at least your highest-risk model with simple extraction prompts before rating it. For the general method behind weeks 5 to 13, see how to identify AI-specific risks, AI risk treatment strategies and monitoring AI risks in production.
“A 90-day programme should produce decisions, not just documentation. By the end, you should know which models need output filtering, which training sets need pseudonymisation, which logs need shorter retention, and who owns each residual risk.” Henri Haenni
A starting point for weeks 5–8
Download our gated AI privacy risk register, pre-structured for AI-specific privacy threats and linked to DPIA and SoA fields.
13. Skills and training paths
No training can make you "27091 certified", because no such certification exists. Choose by whether you build a management system, audit one or govern AI more broadly.
Path: ISO/IEC 42001 Lead Implementer
Path: ISO/IEC 42001 Lead Auditor
Path: ISO/IEC 27701 (2025) Lead Implementer / Lead Auditor
Path: IAPP AIGP (Artificial Intelligence Governance Professional)
Path: PECB AI credentials (for example Lead AI Risk Manager, Certified EU AI Governance Professional, Certified AI Security Professional)
How 27091 fits. 27091 knowledge is a specialist add-on to any of these paths. Short, focused training on 27090 and 27091 typically leads to a certificate of attendance and CPD hours, not a certification.
Course pages: ISO 42001 Lead Implementer · ISO 42001 Lead Auditor · ISO 27701 Lead Implementer · Lead AI Risk Manager
Changelog
- 2026-11-18: Status verified at stage 50.20 (FDIS ballot initiated 2026-11-02). Publication estimate: Feb 2027 (range Jan–Apr 2027).
Copyright notice: this guide quotes only ISO's public title and abstract, with attribution. All other content is our own explanation. For the authoritative text, buy ISO/IEC 27091 from ISO or your national standards body once it is published.




