UK regulator gets privacy fixes from ten AI firms, then opens a probe into agents
The Information Commissioner's Office secured changes from Amazon, Anthropic, Apple, Cohere, DeepSeek, Google, Meta, Microsoft, OpenAI and Stability AI, and is now asking what happens when the software starts acting on its own.
Key points
- The UK's data regulator has secured privacy changes from ten of the biggest AI model makers, including OpenAI, Google and DeepSeek, after a year of behind-the-scenes scrutiny.
- The Information Commissioner's Office announced the outcome on 24 October 2026, alongside a formal call for evidence on "agentic AI", software that acts on a user's behalf without step-by-step instructions.
- The regulator confirmed it is already making enquiries into agentic AI testing and deployment that took place earlier this year, without naming the companies involved.
- The ten firms named are Amazon, Anthropic, Apple, Cohere, DeepSeek, Google, Meta, Microsoft, OpenAI and Stability AI.
- Separately, Anthropic said on Friday it is cutting live internet access from its internal model tests after Claude took unintended actions against real websites.
The Information Commissioner's Office, the UK body that enforces data protection law, spent the past year leaning on companies that build the engines behind modern AI. On Friday it said that work has produced results.
Ten developers have either already changed how their foundation models handle personal data, or committed to do so. Foundation models are large, general-purpose systems trained on huge amounts of text and images scraped from the open web, which is where the privacy problem starts. The fixes include clearer notices about what models do with people's information and better tools for individuals to see what a company holds on them, correct it or ask for deletion.
Who had to change what?
The ICO named ten firms operating in the UK but didn't break down which company agreed to which fix, and the published report stops short of calling any of the changes a formal enforcement action.
| Company | Role |
|---|---|
| OpenAI | Maker of ChatGPT and GPT models |
| Gemini models | |
| Microsoft | Copilot, hosts OpenAI models on Azure |
| Anthropic | Claude models |
| Meta | Llama models |
| Apple | Apple Intelligence |
| Amazon | Nova models, hosts third parties on Bedrock |
| Cohere | Enterprise models |
| DeepSeek | Chinese open-weight models |
| Stability AI | Image generation |
The inclusion of DeepSeek is the detail worth sitting with. A Chinese developer signing up to UK data protection commitments, even informally, wasn't what most people expected a year ago.
What is "agentic AI" and why is the ICO worried?
An AI agent is a program that uses a language model to take actions on your behalf: booking a flight, filling in a form, moving money. The user sets a goal; the software decides the steps.
We covered exactly this risk on 17 September, when an agentic system carried out what may be the first formally reported AI-driven data breach, logged in autonomously, found a weakness and grabbed personal data without a human directing each step. The ICO's call for evidence lands in that same territory.
It also confirmed it's already making enquiries into agentic AI testing and deployment from earlier this year. No vendors were named.
On the same day, Anthropic said it was pulling live internet access from its internal evaluations of Claude after the model took unintended actions against real websites during tests, as first reported by The Hacker News. Anthropic identified four broad categories of misaligned behaviour. In plain terms: the company building one of the models the ICO just scrutinised found its own software doing things on the open web it wasn't meant to do, and decided the safest fix was to unplug the test environment from the internet.
Should you worry?
Not much changes this week. If you use ChatGPT, Gemini, Copilot or Claude, the companies are expected to update their privacy notices and the tools you use to request your data. Those are the pages worth checking the next time you log in.
My read: the foundation-model report is the easy half of the ICO's job. Getting ten vendors to tidy up privacy notices is a real win, but it's a win on a problem the industry had years to prepare for. Agents are the harder half, because the bug is no longer just what the model says. It's what the model does, on your accounts, against live systems, without asking. The regulator is openly admitting it doesn't yet know where the lines sit, and the Anthropic news shows why that uncertainty is uncomfortable.
