When an AI system leaks personal data, the controller notifies that breach to the supervisory authority without undue delay and, where feasible, within 72 hours, unless the breach is unlikely to result in a risk to the people concerned (Article 33 GDPR). Where the breach is likely to result in a high risk, the controller also informs the data subjects themselves, without undue delay and in plain language (Article 34 GDPR). Every breach, notified or not, is documented internally. An AI vendor acting as processor reports a breach to its customer without undue delay, and the customer notifies the authority. If the system is a high-risk AI system, the same event can also be a serious incident that the provider reports under Article 73 of the AI Act. And if the organisation falls under NIS2, as implemented in the Netherlands by the Dutch Cybersecurity Act (Cyberbeveiligingswet), a further route applies, with an early warning within 24 hours. Three notifications, three recipients and three different clocks.
Picture this: you are the privacy officer of a Dutch municipality. On Monday morning a resident reports that the chatbot on the website gave him an answer containing the name, address and the status of the welfare benefit application of someone else. The vendor explains that an update on Friday evening mixed up the conversations of different users. Do you have to notify? Whom? And when does the 72-hour period start, now that the vendor may have known earlier? This post walks through the steps.
In short
A personal data breach is any breach of security leading to the accidental or unlawful destruction, loss, alteration, unauthorised disclosure of, or access to, personal data (Article 4(12) GDPR). The controller notifies a breach to the supervisory authority where feasible within 72 hours of becoming aware of it, unless a risk is unlikely (Article 33(1)). A processor notifies the controller without undue delay (paragraph 2). The notification may be made in phases (paragraph 4), and every breach is documented (paragraph 5). Where a high risk is likely, the data subjects are informed as well (Article 34(1)), except in three cases (paragraph 3). Infringements carry a fine of up to 10 million euros or, for an undertaking, 2 percent of worldwide annual turnover if that is higher (Article 83(4)(a) GDPR). A serious incident with a high-risk AI system is also reported by the provider to the market surveillance authority (Article 73 AI Act), a route that for Annex III systems only really starts to operate from 2 December 2027.
When is it a data breach with AI?
The GDPR looks at the effect on the data, not at the cause. A personal data breach is a breach of security leading to the accidental or unlawful destruction, loss, alteration, unauthorised disclosure of, or access to, personal data (Article 4(12) GDPR). In its Guidelines 9/2022 the European Data Protection Board (EDPB) distinguishes three types: a confidentiality breach (data reaches someone who should not see it), an integrity breach (data is altered without authorisation) and an availability breach (data is lost or can no longer be accessed), possibly in combination (paragraphs 17 and 18). No hack is needed. A software bug or an employee's mistake is enough.
| AI situation | Breach? | Why |
|---|---|---|
| Prompt injection: malicious input makes a chatbot or AI agent reveal data from the documents or systems it can reach | Yes, if personal data is involved | Unauthorised access to and disclosure of data: a confidentiality breach |
| A chatbot shows one user's data to another user | Yes | Unauthorised disclosure. That it happened by accident makes no difference: the definition expressly covers accidental breaches |
| Employees paste customer data into a public AI service with which the organisation has no arrangements | As a rule, yes | The data reaches a party that should not have had access to it. How serious it is depends on which data it was and what the service does with the input, such as storing it or using it for training |
| A leak at the AI vendor or its hosting provider | Yes, for the data it processes for you | The leak affects data you are responsible for. The vendor reports it to you (Article 33(2) GDPR) and you assess whether to notify |
| An AI agent deletes files or overwrites data with incorrect values | Yes | An availability or integrity breach. According to the EDPB, permanent loss or destruction is always an availability breach (paragraph 19) |
With AI it pays to read the definition broadly. A language model that can reach documents, mailboxes or case management systems discloses data to anyone who asks it the right question, even if the security of the underlying system is sound. If the application uses the rights of a general service account instead of those of the user, every successful prompt injection is a possible confidentiality breach.
Notifying the supervisory authority within 72 hours (Article 33 GDPR)
The controller notifies a breach without undue delay and, where feasible, not later than 72 hours after having become aware of it, to the competent supervisory authority (Article 33(1) GDPR). For Dutch organisations that is the Autoriteit Persoonsgegevens, the Dutch Data Protection Authority (Article 6(2) UAVG). If the notification comes later, it must be accompanied by reasons for the delay. No notification is needed where the breach is unlikely to result in a risk to the rights and freedoms of natural persons. According to the EDPB, that can be the case for data that was already public, or for properly encrypted data whose key was not compromised and of which a copy exists (Guidelines 9/2022, paragraphs 75 and 78).
When does the clock start? The EDPB counts from the moment the controller has a reasonable degree of certainty that a security incident has occurred that has led to personal data being compromised (paragraph 31). After a first alert, such as a resident's report, the controller may undertake a short investigation to establish whether a breach has in fact occurred; during that short investigation it is not yet regarded as aware, but the investigation must begin as soon as possible (paragraph 34). If you leave an alert unattended, you cannot later rely on the short investigation.
What goes into the notification? Article 33(3) lists four items.
| Article 33(3) | What the GDPR requires | What you fill in for an AI breach |
|---|---|---|
| (a) | The nature of the breach, where possible with the categories and approximate number of data subjects and records | What happened (for instance user conversations getting mixed up), which data the application showed or revealed, and how many people are affected. You take that number from the application's logs |
| (b) | Name and contact details of the data protection officer or another contact point | The DPO or the privacy team |
| (c) | The likely consequences | For instance identity fraud, phishing that exploits knowledge of a pending application, or harm from exposure of financial or health data |
| (d) | The measures taken or proposed, including to mitigate adverse effects | The application taken offline or rolled back to an earlier version, access rights restricted, the vendor asked for a root cause analysis, data subjects informed |
Notifying in phases is allowed. Where it is not possible to provide all information at the same time, it may be provided in phases without undue further delay (Article 33(4) GDPR). With AI that is the rule rather than the exception: which conversations or documents were affected often only becomes clear after days of log analysis. Make a first notification within 72 hours with what you know and supplement it. The GDPR allows approximations of the number of data subjects, and the EDPB calls notification in phases a safe way to meet the obligation when the extent is not yet known (paragraphs 52 and 53).
The vendor as processor. If the AI vendor processes the data on your behalf, it notifies you without undue delay after becoming aware of a breach (Article 33(2) GDPR). It does not need to assess the risk itself: you do. According to the EDPB, the controller should in principle be considered aware once the processor has informed it, and the EDPB recommends that processors notify promptly and provide further details in phases (paragraphs 44 and 45). The processing agreement must provide that the processor assists you with the obligations of Articles 32 to 36 (Article 28(3)(f) GDPR), and according to the EDPB it can also set an early notification deadline (paragraph 46). So put a concrete deadline in hours in that agreement, together with the information the vendor must supply. If the vendor itself uses a provider of the language model or a cloud platform, a breach at that sub-processor reaches you through the vendor.
Documenting every breach internally (Article 33(5))
The controller documents all personal data breaches, with the facts, the effects and the remedial action taken, so that the supervisory authority can verify compliance (Article 33(5) GDPR). That includes breaches you do not need to notify; the EDPB encourages an internal register of breaches, recording the causes, what took place and the personal data affected (paragraphs 121 to 123). For each breach, record the processing activity it belongs to as it appears in your records of processing activities under Article 30 GDPR. If you change your security after a breach, also update the description of the security measures in those records.
Informing data subjects when the risk is high (Article 34 GDPR)
Where the breach is likely to result in a high risk to the rights and freedoms of natural persons, the controller communicates it to the data subjects without undue delay (Article 34(1) GDPR). The communication describes in clear and plain language the nature of the breach and contains at least the contact details of the DPO or another contact point, the likely consequences and the measures taken (paragraph 2, referring to Article 33(3)(b), (c) and (d)). The EDPB asks for a dedicated message, not one sent along with regular updates or newsletters (Guidelines 9/2022, paragraph 89).
To decide whether the risk is high, the EDPB lists a number of factors: the type of breach, the nature, sensitivity and volume of the data, ease of identification of individuals, the severity and permanence of the consequences, special characteristics of the individuals such as children or other vulnerable people, special characteristics of the controller, and the number of affected individuals (paragraphs 105 to 118). Health data, identity documents and financial data can, according to the EDPB, each cause harm on their own, and in combination enable identity theft (paragraph 108).
The communication is not required in three cases (Article 34(3) GDPR):
- The data was unintelligible to unauthorised persons. The controller had implemented appropriate technical and organisational protection measures, applied to the affected data, such as encryption (point (a)).
- The high risk has been removed. The controller has taken subsequent measures which ensure that the high risk is no longer likely to materialise (point (b)).
- Informing individuals would involve disproportionate effort. In that case there is a public communication or a similar measure informing data subjects in an equally effective manner (point (c)).
If the controller has not informed the data subjects, the supervisory authority may require it to do so, or decide that one of the three conditions is met (Article 34(4) GDPR).
How far the circle of people to inform reaches is shown by an example in the EDPB Guidelines 01/2021. In an attack on a bank's website, data of around 100,000 customers was leaked, and for around 2,000 of them the attacker was also able to log in. The bank only informed the customers whose accounts were compromised. According to the EDPB it should have informed all of the roughly 100,000 data subjects, including those whose accounts were not taken over (Case No. 07, paragraphs 64 to 68).
What the Court of Justice says about security and damage
A breach does not automatically mean that security fell short. C-340/21, Natsionalna agentsia za prihodite (14 December 2023, ECLI:EU:C:2023:986) concerned a cyberattack on the Bulgarian tax agency, after which data of more than 6 million people appeared on the internet. The Court held that unauthorised disclosure or access by a third party is not sufficient in itself to conclude that the measures were not appropriate, that in an action for damages the controller bears the burden of proving that its measures were appropriate, and that a data subject's fear of misuse of his or her data can in itself constitute non-material damage (operative part, points 1, 3 and 6).
In C-687/21, MediaMarktSaturn (25 January 2024) employees handed documents containing personal data to another customer by mistake, who returned them within half an hour. Here too, the employees' error was not in itself enough to find the measures inappropriate. And where it was established that the third party did not become aware of the data, the mere fear of future dissemination is not non-material damage (operative part, points 1 and 5).
For AI this means: keep the evidence of your security measures, such as prompt injection tests, the set-up of access rights and the arrangements with the vendor. In a damages case you must be able to show that they were appropriate. And logs showing who actually got to see which data help determine whether there is damage at all.
Fines
An infringement of Article 33 or 34 falls under Article 83(4)(a) GDPR: a fine of up to 10 million euros or, for an undertaking, up to 2 percent of the total worldwide annual turnover of the preceding financial year, whichever is higher (Article 83). Whether a public authority can be fined is a matter of national law (Article 83(7) GDPR); how that works out for Dutch municipalities is set out in fines under the AI Act and the GDPR.
The link with a serious incident under Article 73 of the AI Act
The AI Act has its own reporting duty, with a different reporter, a different recipient and different deadlines. Providers of high-risk AI systems report serious incidents to the market surveillance authorities of the Member States where the incident occurred (Article 73(1) AI Act). A serious incident is an incident or malfunctioning of an AI system that directly or indirectly leads to the death of a person or serious harm to a person's health, a serious and irreversible disruption of the management or operation of critical infrastructure, the infringement of obligations under Union law intended to protect fundamental rights, or serious harm to property or the environment (Article 3(49) AI Act).
A data breach can fall under the third limb. The GDPR protects the fundamental rights of natural persons, in particular their right to the protection of personal data (Article 1(2) GDPR). If an incident or malfunctioning of a high-risk system leads to an infringement of obligations under the GDPR, that can therefore be a serious incident. A breach does not in itself prove that security fell short, and so is not automatically an infringement of obligations (C-340/21 and C-687/21, below). Not every breach is one: it has to involve a high-risk system and a breach that results from the incident or malfunctioning of that system. When the market surveillance authority receives a report about an infringement of fundamental rights, it informs the national public authorities that protect fundamental rights (Article 73(7) AI Act). In autumn 2025 the Commission published a draft guidance document and reporting template; what they contain is covered in our post on the incident reporting consultation.
The deadlines. The provider reports immediately after establishing a causal link between the AI system and the incident, or the reasonable likelihood of such a link, and in any event not later than 15 days after the provider or the deployer becomes aware of the incident (Article 73(2) AI Act). For a widespread infringement or a serious and irreversible disruption of critical infrastructure the limit is two days (paragraph 3). A widespread infringement includes an act or omission contrary to Union law protecting the interest of individuals that has harmed or is likely to harm the collective interests of individuals residing in at least two Member States other than the one where the act took place or where the provider or deployer is established (Article 3(61) AI Act); a large cross-border breach through a high-risk system can fall under it, and in the event of a death ten days (paragraph 4). An initial, incomplete report is allowed (paragraph 5). For Annex III systems of providers already subject to equivalent reporting obligations under other Union law, reporting is limited to infringements of fundamental rights (paragraph 9).
The role of the deployer. The organisation using the system is, under the GDPR, usually the controller for a breach and then notifies the authority itself. Under the AI Act, the provider reports. Where a deployer identifies a serious incident, it immediately informs first the provider, and then the importer or distributor and the market surveillance authorities; if it cannot reach the provider, Article 73 AI Act applies to it mutatis mutandis (Article 26(5) AI Act). A single event can therefore lead to two reports by two different parties.
From when? Article 73 AI Act sits in Chapter IX and, under Article 113 AI Act, has applied since 2 August 2026. The duty, however, rests on providers of high-risk AI systems, and the rules determining which systems those are (the classification rules and obligations of Chapter III, Sections 1 to 3, including Article 6 and Article 26 AI Act) only apply from 2 December 2027 for Annex III systems and from 2 August 2028 for Annex I systems (Article 113, third paragraph, point (c), AI Act, as amended by Regulation (EU) 2026/1744). In practice the reporting duty therefore only gains an object from those dates. For high-risk systems placed on the market or put into service before that date, the high-risk obligations, including Article 73 AI Act, only apply if they are subsequently subject to significant changes in their design, and systems intended for use by public authorities must comply by 2 August 2030 at the latest (Article 111(2) AI Act). The GDPR notification duty applies now, to every AI application.
Besides this reporting duty, the AI Act also has complaint and whistleblower channels at the AI Office; those are something else than an incident report, see the three reporting routes of AI Act enforcement.
NIS2, the Dutch Cybersecurity Act and DORA
A data breach is often also a cyber incident. Essential and important entities under NIS2 report every significant incident: one that has caused or is capable of causing severe operational disruption or financial loss, or that has affected or is capable of affecting others by causing considerable material or non-material damage (Article 23(1) NIS2, with the threshold in paragraph 3). In the Netherlands this is implemented by the Cyberbeveiligingswet, the Dutch Cybersecurity Act, in force since 15 August 2026 (Article 35 of the Cyberbeveiligingsbesluit). The entity gives an early warning without delay or, if that is not possible, within 24 hours, submits a notification without delay or within 72 hours, and files a final report at the latest one month after that notification, each time to its CSIRT and its competent authority (Articles 26, 27 and 29 of the Dutch Cybersecurity Act).
Two points make the overlap with the GDPR more manageable. Where a NIS2 authority becomes aware in the course of supervision or enforcement that an infringement of the NIS2 obligations may entail a personal data breach that has to be notified under the GDPR, it informs the data protection authority; if that authority imposes a GDPR fine for the same conduct, no NIS2 fine follows (Article 35 NIS2, first and second paragraphs). But the notifications themselves do not replace each other: the GDPR notification and the NIS2 notification are two separate duties.
For banks, insurers and other financial entities, DORA applies to ICT incidents. They report major ICT-related incidents to their competent financial authority (Article 19(1) DORA). According to the NIS2 Directive, DORA is a sector-specific act for financial entities that applies instead on this point (Recital 28 and Article 4 NIS2; in Dutch law Article 31 of the Dutch Cybersecurity Act). Here too, the GDPR notification to the data protection authority still stands.
| Regime | Who reports | To whom | Deadline | Threshold |
|---|---|---|---|---|
| Article 33 GDPR | Controller (the processor reports to the controller) | Supervisory authority (in the Netherlands: the AP) | Without undue delay, where feasible within 72 hours | Unless a risk is unlikely |
| Article 34 GDPR | Controller | Data subjects | Without undue delay | High risk likely |
| Article 73 AI Act | Provider of a high-risk system (the deployer informs the provider first) | Market surveillance authority | No later than 15 days; 2 days for a widespread infringement or critical infrastructure; 10 days in the event of a death | Serious incident (Article 3(49) AI Act) |
| Dutch Cybersecurity Act (NIS2) | Essential or important entity | CSIRT and competent authority | Early warning within 24 hours, notification within 72 hours, final report within one month | Significant incident |
| DORA | Financial entity | Competent financial authority | Initial notification, intermediate report and final report, within deadlines set in implementing rules (Article 19(4) DORA) | Major ICT-related incident |
Three examples
The examples are fictional and serve to make the steps concrete.
Municipality. Back to the chatbot. The resident's report is the first alert; after a short investigation on Monday morning it is established that conversations were mixed up. From that moment the 72 hours run. The logs show that 40 residents received an answer with someone else's data over the weekend, including information about a welfare benefit application. The vendor noticed the bug on Saturday but only reported it on Monday. For the municipality the period starts when it becomes aware itself, but the vendor has not met its duty to report without undue delay (Article 33(2) GDPR); that belongs in the conversation about the processing agreement. Because income data and a pending application are involved, a high risk is likely: the municipality notifies the AP, informs the residents whose data was shown with a dedicated message, and asks the residents who saw the data to delete it. An information chatbot is not an Annex III high-risk system, so Article 73 AI Act does not come into play.
Bank. A bank runs a customer chatbot that looks up transactions through a general service account and has to check for each question whether it concerns the logged-in customer. An attacker discovers that this check falls short and that a cleverly worded question makes the chatbot display overviews of other customers, and in this way retrieves data of 300 customers. This is a confidentiality breach involving financial data. According to the EDPB such data can cause harm on its own (Guidelines 9/2022, paragraph 108), and in the comparable bank case No. 07 the EDPB assessed the breach as high risk (Guidelines 01/2021, paragraphs 64 to 66). The bank notifies the AP within 72 hours and informs all customers whose data may have been retrieved, not only those whose data the attacker is known to have seen; EDPB case No. 07 above shows why. The bank also assesses whether this is a major ICT-related incident under DORA (Article 19 DORA); that report goes to the financial supervisor and does not replace the GDPR notification. A customer chatbot does not fall under Annex III, so a report under Article 73 AI Act is not at issue.
HR. An employer uses a vendor's platform that ranks job applications with AI. An AI system intended for the recruitment or selection of natural persons, in particular to analyse and filter job applications and to evaluate candidates, is listed in Annex III point 4(a) AI Act and is a high-risk system, unless the derogation in Article 6(3) AI Act applies. Because of a misconfigured storage bucket, the CVs and cover letters of 2,000 applicants were publicly accessible for three weeks. Under the GDPR the vendor is a processor and reports the breach to the employer without undue delay; the employer notifies the AP. With contact details, work history and sometimes health or nationality data in a cover letter, a high risk is likely and the applicants are informed; the EDPB reached the same outcome for a leak of job application forms at an employment agency, there caused by a hack (Guidelines 01/2021, Case No. 05, paragraph 55). From 2 December 2027 the vendor, as provider, also assesses whether this is a serious incident within the meaning of Article 3(49)(c) AI Act and, if so, reports it within the deadlines of Article 73 AI Act. That is not self-evident here: Article 73(2) AI Act requires a causal link between the AI system and the incident, and a misconfigured storage bucket is not necessarily an incident or malfunctioning of the AI system itself. If the platform was already in use before that date, the high-risk obligations only apply after a significant change (Article 111(2) AI Act).
Seven steps to be ready for an AI data breach
- Decide in advance who decides. Who assesses, in an incident with an AI application, whether there is a breach, who notifies the authority and who informs data subjects? Involve the DPO, the security function and the owner of the application.
- Cover the vendor in the processing agreement. A concrete reporting deadline in hours, which information it supplies, how it deals with sub-processors such as the provider of the language model, and a fixed contact point (Article 28(3)(f) and Article 33(2) GDPR).
- Log what you need to notify. Which user saw which answer, which sources the application consulted and with which rights. Without those logs you cannot estimate the number of data subjects (Article 33(3)(a) GDPR).
- Investigate an alert straight away. A short investigation is allowed, but the clock runs as soon as you know with reasonable certainty that personal data was compromised.
- Assess the risk and notify. Use the EDPB factors. Notify the authority within 72 hours, in phases if necessary, and inform data subjects without undue delay if the risk is high.
- Check the other routes. If a high-risk system is involved, inform the provider with Article 73 AI Act in mind. If you fall under the Dutch Cybersecurity Act or DORA, that report applies too, with its own deadlines.
- Document and improve. Record the breach in the internal register (Article 33(5) GDPR), update the security measures and the records of processing activities, and review the DPIA if the risk of the processing has changed. When a DPIA is required for AI is explained in DPIA for AI systems: when it is required.
Want to know which GDPR and AI Act duties come into play for a specific AI application? With the case check you describe the application in a few questions.
Frequently asked questions about a data breach with AI
Sources
Newsletter
Every Tuesday, the AI Act week ahead in 5 minutes
A practical briefing on deadlines, new guidance and enforcement, so you know what matters this week. No spam and you can unsubscribe in one click.
Practical and short · No spam · One-click unsubscribe