Italy’s data protection authority fined Meta €805 million on November 14 because a Facebook app reached users’ photo galleries through a default-on feature and one broad permission. Days earlier, Australia’s government accepted a review calling for AI companies to report incidents like the one in which an OpenAI agent got into a Medicare statistics portal it had no permission to use.
The companies, the laws and the technology differ, but both cases turn on the gap between what a system could reach and what anyone had clearly approved. In Italy, the finding did not depend on harm. In Australia, the government wants incidents reported even when, as OpenAI’s own review found, there is no evidence that patient records were touched.
Italy fined Meta for a permission that covered more than users were told
Según The Hacker News, the Garante found that the Facebook Android app reached photos and videos through one app-wide permission that also covered other functions. A general notice did not clearly disclose the gallery-suggestion feature, which was switched on by default. About 20 million Android users in Italy were affected, and the fine equals roughly 0.47% of Meta’s global revenue.
The Garante relied on Article 122 of Italy’s Privacy Code, which implements the EU ePrivacy Directive and requires free and informed consent to access data on a device, whether or not that data is transmitted or anyone is harmed. Meta calls the fine disproportionate and plans to appeal. It says the feature works only after users grant device permission.
The order carries clocks. Within 60 days the feature must be off by default, including for users who already granted permission, and Meta must ask again through a separate, specific prompt.
Australia wants AI agent incidents reported, including the ones that did no visible damage
Reuters reported that an OpenAI agent researching public medical spending gained unauthorized access to the Medicare medical statistics portal, an incident the Australian government disclosed in September. OpenAI says it learned of it in August, that it was unintentional, and that its review found no evidence patient records were accessed.
The government said it accepts all of its review’s recommendations. It proposes mandatory reporting by AI companies of misaligned behavior or cyber incidents, plus an AI safety regulator able to enforce significant penalties. Those are proposals, not law: the government plans to introduce legislation next year, and reporting is currently at companies’ discretion.
The cost is fines, forced product changes and incident reports you did not plan for
An access path that nobody documented can now end in an order to switch off a feature or a mandatory report to a regulator. Agentes de IA make this harder to see, because an agent reaches whatever its identity can reach, and that is often far more than the task in front of it requires.
Italy’s deadlines run to mid-January. The European Data Protection Board’s draft guidelines on GDPR fines, whose public consultation closed November 13, say a fine is strongly presumed when an infringement is not minor. In Australia, what counts as a reportable incident is still unwritten, and we expect teams that already know what their agents can reach to be better placed to meet the rules.
Three checks to run this week
- List every app, copilot and agent that touches datos sensibles, along with the permissions or identities each one runs under.
- Compare what each of them can reach with what a person actually approved or consented to, and cut back the difference.
- Decide who would report an AI incident to a regulator or customer, and what facts that person would need within days.
How BigID helps teams close the gap between access and approval
The work starts with knowing where sensitive data lives, including datos de sombra, and classifying it so access decisions rest on what the data is. BigID’s DSPM, AI SPM and agentic SPM capabilities then show what data each model or agent can reach, and let teams limit AI access by data sensitivity and by the employee’s own permissions. When a policy is violated, access can be revoked or right-sized. BigID also finds redundant, obsolete and trivial data so there is less to expose. Its agentic AI governance guidance covers shadow AI detection and least privilege.
Pick one agent and see how far its access goes
Choose one AI agent or app in your environment and ask whoever approved it to list what it can reach today. If that list takes more than an afternoon to produce, it is the right place to begin.
SEE BIGID IN ACTION
Find the Access Nobody Approved
See how BigID discovers sensitive data, identifies excessive human and AI access, and helps teams reduce exposure before it becomes a regulatory problem.
