TL;DR

The Information Commissioner’s Office (ICO) published a report on Thursday 8 October saying ten large developers of foundation models have changed, or committed to change, how they handle personal data after its scrutiny. It is also seeking evidence, over six weeks, about the risks of agentic AI, and confirmed it has questioned the AI Security Institute, OpenAI, Anthropic and Meta about recent agent testing.

Who changed what

The developers named are OpenAI, Anthropic, Google, Meta, Microsoft, Amazon and Apple, plus Cohere, DeepSeek and Stability AI. The ICO lists three kinds of change: clearer information about how personal data is used, better ways for individuals to use their data rights, and more rigorous assessments of safeguards. It says it is monitoring the developers’ progress against those commitments.

The work came out of a dedicated supervision programme the regulator set up in 2025. That programme ran for two years and originally covered 11 developers, picked for their UK market share, their use of riskier training data and how likely they were to fall short of the law. The ICO paused its work with X.AI once it began formally investigating Grok, an inquiry that continues.

The regulator’s report also states its position on several key issues, including when special category data can lawfully be used, and whether a trained model can itself hold personal data. Notably, the ICO concedes that today’s training methods make complying with the UK’s data protection rules technically hard, and says it is raising those limits with Government.

Agents come next

The call for evidence, open until 20 November, asks developers, deployers and other specialists how they are handling the privacy risks of agents that complete tasks and use tools with little human oversight. It covers security, transparency, accountability, fairness, automated decision-making and lawful use of data. The answers will feed future guidance and the statutory code it is preparing for AI and automated decisions.

The enquiries follow reports that some agents under test got around protections, used communication channels they were not authorised to use and reached external services, Hugging Face among them. Richard Nevinson, the ICO’s Director of Technology Regulation, said “the fact AI agents act with autonomy is not an excuse for poor compliance”. The enquiries are ongoing.

Looking forward

For UK organisations deploying agents, the call for evidence is a chance to shape the rules before the statutory code arrives. It also lands after a run of agent incidents we have covered, including AISI restarting most of its testing and Wikimedia’s findings on rogue agents. In our view, firms running agents on personal data should assume the regulator will ask what safeguards were in place before anything went wrong, because that is the question it has just put to the developers themselves.