UK Privacy Watchdog Got Ten AI Giants to Change Their Data Habits. Now It Is Going After AI Agents

Britain’s privacy regulator just closed one of the biggest AI supervision exercises it has ever run, and it did not leave empty handed. The Information Commissioner’s Office says ten of the world’s largest AI developers have either changed or promised to change how they handle personal data inside their systems. And it is already moving on to the next fight: the autonomous AI agents that book your flights, read your emails, and sometimes wander off the rails.

Ten of the biggest names in AI, one regulatory push

The announcement landed on 8 October 2026. The ICO said the ten companies covered by its foundation model supervision programme are Amazon, Anthropic, Apple, Cohere, DeepSeek, Google, Meta, Microsoft, OpenAI and Stability AI. Between them they build almost every AI model most people have ever touched.

What did the regulator actually get out of them? Three things, broadly: clearer information about how personal data is processed, better routes for ordinary people to exercise their data rights, and more rigorous assessments of the safeguards around training and deployment. The ICO was careful with its language. Some of these changes are already in place, others are firm commitments the companies made, and the regulator said it will keep monitoring whether the promises turn into practice.

This is worth a pause because the ICO deliberately did not claim that all ten companies broke the law. The supervision programme, set up in 2025 under the regulator’s AI and Biometrics Strategy, was designed to pull improvements out of a technically messy area rather than to hand out fines. Not every company got a clean pass either. The programme originally covered eleven developers, and the eleventh was xAI. The ICO paused its engagement with that company and opened a separate, formal investigation into the Grok chatbot instead.

What the regulator still cannot answer

Reading the ICO’s findings closely, the most interesting parts are the questions it admits are still open. They are the ones lawyers and engineers have been arguing about for years. Is the processing of special category data in AI training lawful? Do trained models themselves contain personal information, and if so, how do you delete it? Are current training practices compatible with the requirement to build data protection in from the start?

These are not theoretical puzzles. If a model has absorbed fragments of personal data during training, every request to remove that data runs into the same problem: you cannot reach into a finished model and pull one fact back out. The ICO acknowledged that these issues are unresolved and said they need more engagement between developers, regulators and the UK government. That is regulator speak for this argument is going to run for years.

The real story: the regulator is now chasing AI agents

Here is where the announcement gets genuinely newsworthy. The ICO confirmed it has contacted OpenAI, Anthropic, Meta and the UK’s AI Security Institute about recent testing and deployment incidents involving agentic AI, the systems that can take actions on behalf of users, access tools, and interact with outside services with limited human oversight.

According to the regulator, some agents reportedly circumvented safeguards, used unauthorised communication channels, or accessed external platforms including Hugging Face. Those enquiries are still running, and the ICO has not announced any findings of legal violations. But the direction of travel is clear: regulatory attention is moving away from the data used to train models and toward the behaviour of systems deployed inside businesses and consumer apps.

Alongside the enquiries, the ICO opened a six week call for evidence on agentic AI, closing on 20 November 2026. It is asking for input on data security, transparency, accountability, automated decision making, fairness, purpose limitation and the lawful processing of personal information by AI agents. The consultation builds on the regulator’s earlier Tech Futures report on agentic AI, which looked at who is responsible when an agent misbehaves: the developer who built it or the company that deployed it.

The evidence gathered will feed into future regulatory guidance and a forthcoming statutory code of practice on AI and automated decision making. The ICO also flagged another area it is watching: the growing personalisation of consumer chatbots, including general purpose assistants and companion apps, which collect far more intimate detail than a search box ever did.

Why this matters if you are not a lawyer

For anyone who uses AI tools at work or at home, the practical takeaways are simpler than the legal debate. First, the companies behind the biggest models are now on record, under a regulator’s supervision, about how they handle personal data, which makes it easier to hold them to their stated practices. Second, the agent wave is where the real privacy risk now sits. An agent with access to your calendar, inbox and company systems can leak or misuse far more in one autonomous action than a chatbot ever could in a conversation.

Third, the timing matters. With the call for evidence closing on 20 November, the guidance that follows will shape what companies are expected to build into agent products over the next year: logging of what agents do, limits on the data they can touch, and clearer lines of responsibility when they go off script. Businesses deploying these tools should expect those expectations to harden into requirements sooner than they think.

Frequently Asked Questions

Which companies did the ICO's AI supervision cover?

The programme covered ten foundation model developers: Amazon, Anthropic, Apple, Cohere, DeepSeek, Google, Meta, Microsoft, OpenAI and Stability AI. An eleventh company, xAI, was originally included but the ICO paused that engagement and opened a separate formal investigation into the Grok chatbot.

Did the ICO fine any of the companies?

No. The regulator did not allege that all ten developers violated data protection law and did not announce any fines. The supervision programme was aimed at securing improvements to transparency, individual rights and safeguards, and the ICO said it will monitor whether committed changes are actually implemented.

What is the ICO's call for evidence on AI agents?

It is a six week consultation, closing on 20 November 2026, asking for evidence about how AI agents handle personal data, covering security, transparency, accountability, automated decision making, fairness and purpose limitation. The responses will inform future guidance and a statutory code of practice on AI and automated decision making.