GDPR Gets an AI-Training Rulebook While the AI Act Clock Moves: What Data Teams Need to Do Now

GDPR Gets an AI-Training Rulebook While the AI Act Clock Moves: What Data Teams Need to Do Now

Week of 2 November 2026 | Compiled by Team Paringa

The rules on how personal data can feed AI are being rewritten in Brussels, while basic data-handling failures keep making headlines. This week's brief covers what changed, and what it means for teams holding PII.

What Happened This Week

🔴 EU states back a GDPR rewrite that keeps "personal data" intact and adds pseudonymisation standards

The Council of the EU agreed its negotiating position on the Digital Omnibus on 21 October, with IAPP reporting it on 22 October, just ahead of this week's window but still the most significant live development. The Council kept the existing GDPR definition of personal data and rejected the Commission's proposed change to Article 4(1). It added an amendment setting out pseudonymisation standards, and proposed new safeguards that would allow sensitive data to be used for biometric verification and AI model training, with a clarified lawful basis for processing personal data in AI training.

The Council also limited cookie banners by naming cases where consent is not required, such as strictly necessary cookies, audience measurement and device security. It opposed a single EU-level cyber incident reporting entry point and proposed one per member state instead. Parliament has not yet settled its own position, so trilogues cannot start. IAPP also reported a LIBE hearing on Omnibus scepticism, and the EDPB and EDPS have previously raised concerns over the personal data definition.

For anyone running AI on customer data, the direction is clear. Personal data stays broadly defined, so tokenised or pseudonymised data is only as safe as the standard it meets, and that standard is about to be written down. Final text is not settled, and the Parliament could push back.

Sources: IAPP, Council of the EU, IAPP

If the Council text survives, defensible pseudonymisation becomes a named standard rather than a matter of interpretation. That favours architectures where raw PII never reaches the model or the data lake, and only tokens do. Do not wait for trilogues to finish. Map where identifiable data enters your AI pipelines today, because the teams that already isolate it in a vault will meet any final standard with the least rework. A short data-flow review is a sensible first step.

🟡 The AI Act delay is law, and national enforcement is arriving

Regulation (EU) 2026/1744, the AI Omnibus, was published in the Official Journal on 24 July and entered into force on 27 July, according to a Mondaq law-firm summary. Stand-alone high-risk systems under Annex III move from 2 August 2026 to 2 December 2027, and high-risk AI embedded in regulated products moves to 2 August 2028. IAPP's April reporting set out the same two dates as the previously agreed postponements.

Not everything slipped. General-purpose AI obligations continue, transparency duties from 2 August 2026 still apply, and two new prohibitions on intimate-imagery and CSAM generation apply from 2 December 2026. The AI Office also gains expanded supervisory powers. National enforcement is now taking shape: Italy's Legislative Decree 179/2026 entered into force on 23 October, naming AgID as notifying authority and ACN as the main market surveillance authority.

Sources: Mondaq, IAPP, IAPP

The extra 16 months for high-risk systems is breathing room, not a reason to pause. Training-data governance, access control and logging are the hard parts of any high-risk programme, and they take longer than the paperwork. Use the window to build them properly once, in a way that also satisfies GDPR.

🟡 Applebee's leaves a database password in its Android app

TechCrunch reported on 30 October that a database credential was embedded in Applebee's Android app, letting anyone who extracted it log in to the customer database. UK researcher Robert Wiggins found it using AI to search the app's source code, and TechCrunch verified the flaw with a dummy account. Exposed fields included names, emails, dates of birth, phone numbers, ZIP codes, veteran status and login IP addresses.

Applebee's would not say how many customers were affected; the company had at least 8.5 million rewards members last year. Parent Dine Brands says it found no evidence of unauthorised access so far, and the credential has been fixed. The researcher could not find any channel for reporting security issues.

Sources: TechCrunch

This is a single-source story with unknown scale, so treat the headline numbers with care. The lesson is structural: when a credential opens a table of raw PII, one mistake exposes everything. Storing sensitive fields as tokens behind a vault limits what any single leaked key can reveal. Also note that AI-assisted code scanning now makes this kind of discovery cheap for attackers as well as researchers.

What Data and Privacy Teams Are Dealing With

Regulation is shifting from "can we use personal data for AI" to "prove how you protected it". The Council text and the AI Act dates both point to documented controls rather than blanket restrictions, but deadlines are moving and positions are unsettled, which makes planning harder.

At the same time, ordinary exposure incidents continue, and ICO-style enforcement for security failings is a live risk for UK firms. The practical response is to reduce the amount of raw PII in circulation, not to wait for final rules.

Three Things Worth Watching Next Week

1. European Parliament position on the Omnibus — Trilogues cannot start until Parliament settles its stance. Watch whether it backs the Council on pseudonymisation and AI training.

2. Applebee's follow-up — Whether Dine Brands confirms customer numbers or regulator notifications will show how the incident is classed.

3. 2 December 2026 prohibitions — The new AI Act bans on intimate-imagery and CSAM generation are five weeks away, and vendors will start announcing compliance.

Could not verify: The Register was unavailable to automated retrieval (robots restrictions), so its coverage was not used; no ICO fine announced this week was found; the Council press release URL was seen in search results but could not be opened, so Council details rely on IAPP.

Questions? hello@paringa.ai

Paringa helps organisations keep sensitive data out of AI pipelines by design, using vault-based tokenisation through Skyflow. We turn regulatory change into a practical architecture plan. | 2 November 2026 | © Paringa Ltd