The ICO Puts Foundation Model Training Data on Notice

Week of 12 October 2026 | Compiled by Team Paringa

Regulators are moving from guidance to expectations on how personal data enters AI systems, and the organisations feeding that data are the ones carrying the risk.

What Happened This Week

🔴 ICO secures commitments from ten AI developers and opens an agentic AI inquiry

The UK Information Commissioner's Office published its report "Building trust and transparency into generative AI development" on 9 October. Ten developers committed to its expectations: Amazon, Anthropic, Apple, Cohere, DeepSeek, Google, Meta, Microsoft, OpenAI and Stability AI. X and xAI are not on the list, and the ICO says it will investigate them for possible personal data violations.

The substance is a four-part test for using personal data to train foundation models. A developer needs an identified lawful basis, meaningful transparency, working mechanisms for people to exercise their rights, and safeguards that materially reduce risk. The ICO maps these to existing UK GDPR, so this is not new law. It is a regulator telling the market how it will read the law it already has.

Alongside the report, the ICO opened an evidence call on agentic AI, closing 20 November 2026, to inform future guidance. It has put questions to OpenAI, Anthropic, Meta and the UK AI Security Institute about recent agent tests and deployments. Agents that act on personal data with minimal human oversight raise sharper questions on purpose limitation, automated decision-making and data minimisation than chat interfaces do.

For enterprises the point is downstream. If the model builders must show lawful basis and risk-reducing safeguards, their customers will be asked the same questions about the data they send in at inference time, in fine-tuning sets and in retrieval stores.

Sources: Biometric Update

The ICO has just defined "safeguards that materially reduce risk" as a live expectation, and tokenising or removing PII before it reaches a model is the most defensible answer to it. Do not wait for the agentic guidance to land. If you run, or plan to run, agents against customer data, map today which fields they can read and replace raw identifiers with tokens. Use the 20 November call as a deadline to get your own position documented. Regulators reward organisations that can show the control, not just the policy.

🟡 44,000+ Article 21 objections hit NHS England's £330m Palantir platform

More than 44,000 citizens have invoked their right to object under Article 21 UK GDPR to the handling of their data on NHS England's Palantir-run platform, according to reporting by The Guardian on 30 September and picked up in the IGS data protection digest of 5 October. The contract is worth £330m over seven years and was signed in late 2023, with a break clause due early next year.

Mass objection is a different risk from a breach. Each objection is a rights request that the controller must assess, and the volume turns a legal right into an operational load. It also puts the break clause under political and legal pressure.

Elsewhere in the same digest, two UK public bodies reported incidents to the ICO: Dyfed-Powys Police after a cyber incident that may have exposed staff data, and an NHS trust that suspended ten employees over inappropriate access to a child's records. Neither is large, but both are the pattern the ICO sees most, which is insiders and access control.

Sources: Information Governance Services digest, 5 October 2026 | The Guardian

Rights requests at this scale are only manageable if you know where an individual's data lives. Vault-and-token architectures make an objection or deletion a single operation rather than a hunt across systems. If your own DSAR process relies on people searching spreadsheets, test it with 100 requests, not one.

What Data and Privacy Leaders Are Dealing With

The AI Act timeline gave teams breathing room but not an exit. The May 2026 provisional agreement on the Digital Omnibus moves standalone high-risk obligations to 2 December 2027 and embedded products to 2 August 2028, while a new ban on non-consensual sexual content generation applies from 2 December 2026 (CMS Law, noting formal endorsement was still to follow). Meanwhile UK enforcement is not waiting. Earlier in 2026 the ICO fined Reddit £14.47m over children's data, and the common thread in recent penalties is stale or missing DPIAs.

The practical squeeze is that AI adoption is outrunning data inventory. Boards want agents in production, regulators want lawful basis and safeguards, and privacy teams are asked to sign off without a map of where personal data flows.

Three Things Worth Watching Next Week

1. ICO follow-up on X and xAI — The ICO has said it will investigate. Watch for scope and any formal notice.

2. Agentic AI evidence submissions — The call closes 20 November. Expect industry and civil society positions to surface in coming weeks.

3. NHS England break clause pressure — With 44,000+ objections on record, watch for official response ahead of the early-2027 break clause.

Questions? hello@paringa.ai

Paringa helps organisations turn privacy regulation into a build-ready architecture, with Skyflow data privacy vault capabilities at the core. We focus on practical controls that satisfy regulators and keep AI projects moving. | 12 October 2026 | © Paringa Ltd