Regulators Fine the Data Path, Not the Model
Regulators are not waiting for a settled theory of AI before they fine the way companies handle data. Italy’s data protection authority, the Garante, fined Meta Platforms Ireland 805 million euros on November 14, 2026, over a Facebook for Android feature that suggested photos and videos from users’ phone galleries. The Garante found that the app reached those photos, videos, and metadata without specific consent, and it said plainly that the proceeding did not concern whether the content was used to train AI.
That detail is the lesson for anyone who owns compliance. An 805 million euro fine had nothing to do with a model. It was about who accessed what data, on what authority, and whether the person involved had meaningfully agreed. A day earlier, state breach notices from Uber had traced an exposure of drivers’ license data to stolen credentials at an AI vendor. Two different events, one common thread. Regulators and the public are following the data path.
For a CISO or a Chief Compliance Officer, the practical question is not what a regulator thinks of your model. It is whether you can produce the evidence for every access to regulated data, including access by AI tools, AI agents, and the vendors behind them. This post walks through what the Garante decided, what the decision left open, and what an evidence package must contain.
Key Takeaways
- Enforcement followed the data, not the model. Italy’s Garante fined Meta 805 million euros over gallery access without specific consent, and noted that the proceeding did not concern AI training.
- Consent is an evidence problem. The Garante found that one app-wide permission covered several functions with different purposes and was preceded by a generic notice.
- Lead-regulator arguments do not always help. The Garante rejected Meta’s one-stop-shop argument because the ePrivacy rules it applied do not provide for that mechanism.
- Vendor chains stay inside your audit scope. Uber’s state notices traced a driver data exposure to compromised credentials at an AI infrastructure provider.
- Audit readiness lags the rulebook. Grant Thornton found that 78 percent of surveyed leaders lack strong confidence they could pass an independent AI governance audit within 90 days.
What the Garante Decided
The Garante’s announcement, dated November 14, 2026, fined Meta Platforms Ireland Limited 805 million euros over a gallery suggestions feature in the Facebook app for Android. The feature had been available since 2023 and proposed photos and videos from a phone’s gallery as ideas for posts. Coverage by The Hacker News adds that the fine equals about 0.47 percent of the group’s global revenue and covers about 20 million Android users in Italy.
The findings are specific. The app requested gallery access through a single permission that covered the whole app and several functions with different purposes. A generic notice about enabling other features came first and did not make clear that users were also authorizing gallery suggestions. The feature was enabled by default. The Garante added that a fully granted permission of this kind lets an app read the whole photo archive on an ongoing basis, and that neither the user nor the authority can verify which information is used.
The legal basis matters as much as the facts. The Garante applied Article 122 of Italy’s Privacy Code, which implements the EU’s ePrivacy Directive, not the GDPR. It treats a smartphone as part of a person’s private sphere, so any access to what the phone stores requires free and informed consent, whether or not the data is transmitted or any harm results.
The order is concrete. Within 60 days, Meta must ask for consent through a separate, specific request that explains the feature’s purpose, the gallery access, and how to turn it off, and it must keep the feature off by default, including for users who already granted the permission. Meta must report to the Garante within 30 days on what it has started and within 90 days on the measures it has adopted. If Meta misses the 60-day deadline, the feature cannot be activated for anyone without valid consent.
Meta plans to appeal and calls the fine disproportionate. It also says the feature works only after users give permission on their device and that users can change that permission in settings. The decision is an administrative order, and the appeal is unresolved, so treat the fine as a signal about enforcement direction and not as settled law.
The AI Question the Decision Left Open
The Garante said the proceeding did not concern whether gallery content was used to train AI models, which Meta denies. It also said it reserves the right to assess separately a newer Facebook feature, launched in 2026, that allows cloud processing of gallery content. So the AI question is not closed. It is deferred.
Even so, the structure of the violation will look familiar to anyone deploying AI assistants inside an enterprise. Consider an internal assistant that receives a single broad connector permission to a shared file store. The permission was approved once, by an administrator, under a generic description such as “enable productivity features.” Later, a team turns on a summarization function and an agent that drafts customer replies. Neither function was described when the permission was granted, and the record of who agreed to what is a ticket from eighteen months ago. That is the pattern the Garante penalized, translated to the enterprise.
The point is not that enterprise AI deployments are illegal. It is that the compliance question for AI features rarely starts with the model. It starts with whether access to data was granted for a specific purpose, by someone entitled to grant it, and whether the record shows it. Teams that treat AI as a model problem tend to build testing programs. Teams that treat it as a data access problem build evidence. Regulators, as this week showed, ask for the evidence.
Consent Is an Evidence Problem
Consent in the Garante’s logic must be specific. A generic notice and a bundled permission did not meet that standard in this case, and the Garante did not need to prove harm. Access alone, without valid consent, was the infringement.
For a compliance team, that shifts the definition of done. A consent screen shown at install is not a control. A control is a record that can answer questions an investigator will ask. Which prompt did the user see, and in which version? What did it say about purpose? What did the user choose, and when? Can the choice be withdrawn, and does the system honor the withdrawal for data already reachable? If the only answer is a screenshot of a design file, the evidence is missing.
The same discipline applies to employees and customers inside an AI workflow. When an employee connects an assistant to a mailbox, or a customer approves a data share with an agent acting on their behalf, the compliance record must show the scope that was approved and the scope that was used. The gap between those two is where findings come from.
Data minimization is the quiet companion to specific consent. Granting a tool access to a whole archive when it needs one folder creates the same unverifiable condition the Garante described, where nobody can say which information is used. Scoping access to the minimum necessary, a principle at the center of data minimization and privacy by design, shrinks both the exposure and the evidence burden.
The Vendor Chain Stays in Your Audit Scope
The Uber incident shows the other half of the data path. TechCrunch reported that Uber was notified on October 9 by a provider that hackers had stolen internal credentials and used them to access Uber’s data. TechCrunch identified the provider as Fireworks AI, a company that lets developers use open AI models. Uber’s spokesperson attributed the notices to a security incident at a third-party vendor, Fireworks.ai, that had access to some Uber data.
The filings give the scale so far. According to notices TechCrunch reviewed, driver’s license information was stolen for 2,567 people in Massachusetts, 3,430 in South Carolina, and 1,873 in Rhode Island, with an Illinois count pending and California drivers also notified. Uber did not say how many people are affected overall. Reuters added that the California notice, dated November 10, did not name the provider or list the data involved, and that Uber said it had identified no misuse of the data to date. Fireworks said separately that it identified unauthorized access to some of its systems that affected customer data.
Two details deserve attention. The exposure sat outside Uber’s own systems, in an account at a vendor that held a copy of Uber’s data. And the record that regulators received was a state breach notice, which means the evidence of what the vendor held and why had to be reconstructed after the fact. State data privacy laws and breach notification statutes often place obligations on the organization that owns the data, and a contract with a vendor does not by itself move that accountability.
Schellman’s 2026 State of AI Governance report suggests how little attention this link gets. In its survey of 525 US professionals, only 36 percent said their boards or leadership had formally discussed third-party or vendor AI risk, and only 69 percent were confident they could govern AI embedded in third-party tools. Schellman calls third-party AI the number-one blind spot in enterprise AI governance, and the firm sells AI assessments and certification, so the emphasis is partly its own. The survey is unweighted and US-only. The direction is still hard to argue with, because the Uber notices describe exactly that gap.
Treat every vendor that receives regulated data as part of your own audit scope. That starts with a register of which data categories each vendor holds, under which credentials, with what rotation and revocation rules. The discipline is the core of third-party risk management and supply chain risk management, and AI vendors, which often hold datasets, billing records, and prompts, belong on the list first.
Audit Readiness Lags the Rulebook
If enforcement follows the data, readiness is the ability to show the data path on demand. Two surveys suggest most organizations are not there, even when they feel prepared.
Grant Thornton’s 2026 AI Impact Survey asked 950 senior business leaders in February and March 2026 whether they could pass an independent AI governance audit within 90 days. Seventy-eight percent lack strong confidence, and 22 percent are very confident. The firm notes that its results reflect perceptions and that its use of the word audit is conceptual. It sells AI advisory services, and its finding that integrated AI correlates with better outcomes is its own interpretation.
Readiness also varies by rulebook. In Schellman’s survey, fielded from April 13 to May 11, 2026, 89 percent of respondents had taken specific actions to prepare for US regulations, 29 percent for the EU AI Act, and 12 percent for Asia-Pacific requirements. Those figures describe actions taken, not outcomes, and they come from a US-only sample. They still show something useful. Preparation follows the regulator a company expects to meet, and a decision like the Italian one comes from a regulator many US-centered programs have not mapped. Oversight shows a similar pattern. Only 54 percent of Schellman’s respondents regularly report AI governance to their board or an equivalent executive body.
The Kiteworks survey work points the same way. Kiteworks Data Security and Compliance Risk: 2026 Annual Survey Report found that 63 percent of its 459 respondents had faced a compliance consequence in the 12 months before the survey, defined as an audit finding, a required remediation plan, a board escalation, a contractual penalty, or a formal regulatory investigation. The survey was conducted in the second quarter of 2026, the findings are self-reported, and Kiteworks ran it, so read the figure as one data point among several.
None of these surveys measures what a regulator will ask. They measure how prepared organizations believe they are. The gap between belief and evidence is where an investigation starts.
What an Evidence Package Must Show
Regulators regulate data, not models. A package that answers an inquiry about AI access to regulated data does the same work as a package for any other access, with the AI-specific fields included.
For each access to regulated data by a person, an AI tool, or an agent, the record should show the following.
- The identity that acted, with an indication of whether it was a person or an agent, and the person on whose behalf an agent acted.
- The data reached, with its classification and the system it came from.
- The purpose and the legal or contractual basis, including the consent or authorization that applied and its version.
- The policy that evaluated the request, and the result.
- The time the access began and ended, and the method by which access was revoked.
- For vendor flows, the vendor, the credentials used, and the notification obligations that apply.
Notice what is not on the list. The model’s architecture, its training data, and its benchmark scores do not appear. Those matter for other reasons, and an auditor may ask about them. The Italian decision and the Uber notices both center on access and authorization.
Producing that package takes a single, consistent audit trail. An organization with one log for email, another for file transfer, a third for AI tools, and a fourth for vendor connections can answer the questions eventually and slowly. An organization with a unified audit log that covers those channels can answer them quickly and with confidence that nothing was missed. The Kiteworks Control Plane is built on that premise. It applies one policy model to the people and the agents that access sensitive content, and Kiteworks Compliant AI records each AI request under that model. Those capabilities do not replace legal analysis or a consent program. They supply the evidence layer that both depend on.
What Compliance Leaders Should Do Next
This section is general information, not legal advice, and the right steps depend on your jurisdictions and your counsel. The sequence below is built to be practical for a compliance team in the next quarter.
- Map every AI feature that touches regulated data to a basis for access. List each assistant, agent, connector, and AI vendor, then record the consent, contract, or legal basis that covers its access and who granted it.
- Replace broad permissions with scoped ones. Wherever a single permission covers several functions with different purposes, split it. Narrow scope shrinks both exposure and the evidence you must produce.
- Make consent specific and versioned. Keep the prompts, the version history, the user choices, and the withdrawal records together, so one query returns the full story.
- Build the vendor register. For each vendor that holds regulated data, record data categories, credentials, rotation practice, and notification terms, and review the register whenever a vendor reports an incident.
- Rehearse an evidence pull. Choose one recent AI access event and time how long it takes to assemble the six items above. Update the incident response plan with what you learn.
- Put readiness on a board schedule. Report readiness as the ability to produce evidence, and track it the same way you track other regulatory compliance obligations.
Start with the rehearsal. Opinions about readiness are cheap and a timed drill is not.
To learn more about building audit-ready evidence for AI access to regulated data, schedule a custom demo today.
Frequently Asked Questions
Yes. The Garante’s decision shows that unauthorized access to personal data can be penalized even when AI training is not at issue. The Garante said the Meta proceeding did not concern whether gallery content was used to train AI models, yet it fined the company 805 million euros for access without specific consent. Treat any AI feature that touches personal data as a data access question first, and review your data privacy controls for it.
You need a versioned record of the separate prompt, what it said about purpose and data access, what the user chose, and when. The Garante ordered Meta to present a separate, specific request that explains the feature’s purpose, the gallery access, and how to disable it. Keep those records together with withdrawal events so one query returns the full history, and store them in an audit log that your auditors can query.
In the Garante’s decision it was not. The Garante found that a single permission covering the whole app and several functions with different purposes, preceded by a generic notice, did not produce specific consent under Article 122 of the Italian Privacy Code. Rules differ by jurisdiction, so confirm the position with counsel, and split broad permissions by purpose as a safe design default. A data protection impact assessment is a useful place to record that decision.
Keep a vendor register that records, for each vendor, the data categories it holds, the credentials used to reach them, the rotation and revocation practice, and the notification terms. The Uber notices show why. Stolen credentials at an AI vendor exposed driver’s license data, and Schellman found that only 36 percent of surveyed organizations had discussed third-party AI risk at the board or leadership level. Feed the register into your vendor risk management process and review it after every vendor incident.
Tell the board the truth in terms of evidence, not confidence. Grant Thornton found that 78 percent of surveyed leaders lack strong confidence they could pass an independent AI governance audit within 90 days, so the gap is common, and Schellman found that only 54 percent regularly report AI governance to their board. Present a plan that names an owner, the six evidence fields, and a date for a timed evidence drill, and route it through your governance, risk, and compliance program.
Additional Resources
- Blog Post The Tug-of-War Over Your Data: How the CLOUD and SHIELD Acts Pit Security vs. Privacy
- Blog Post Secure Sensitive Data by Mapping DSPM to Your Compliance Goals
- Brief Top 3 FERPA Violations and How to Avoid Them
- Blog Post Executive Order 14117: Protecting Americans’ Bulk Sensitive Personal Data
- Blog Post Need NIS2 Compliance? Start With ISO 27001