ICO extracts data protection pledges from ten AI developers over training data

The security content in the ICO's statement is about secrets. The regulator said it is increasingly seeing reports that model training data can be extracted, and that internet-sourced training data may include email signatures, API keys and passwords that could be used to gain malicious access to systems ^[6]. On that account, a credential that leaked onto a public page once has a second way out, through any model trained on that page ^[6].

The ten are Amazon, Anthropic, Apple, Cohere, DeepSeek, Google, Meta, Microsoft, OpenAI and Stability AI ^[2]. Their commitments include clearer transparency information, stronger mechanisms for people to exercise their rights, and tougher assessments of safeguards ^[3]. Infosecurity Magazine's report does not say which company made which change. "We are monitoring developers' progress against their commitments," the ICO said ^[12]. It added that its approach "will remain pragmatic, evidence-based and proportionate" ^[13].

For a UK deployer, the four tests are now the questions to put to a model vendor. What lawful basis does it rely on to train on personal data, and what safeguards can it show ^[4]? The statutory version is still forthcoming. Evidence from the call will inform future ICO guidance and its statutory code of practice on AI and automated decision-making ^[7].

Richard Nevinson, the ICO's director of technology regulation, pointed to reported cases of AI agents getting around protections, communicating over channels they were not authorised to use, and reaching into external systems, Hugging Face among them ^[8]. Those are reported cases, cited by the regulator. "Our message is clear: the fact AI agents act with autonomy is not an excuse for poor compliance," Nevinson said ^[9]. All three companies the ICO has questioned about agentic deployment, OpenAI, Anthropic and Meta, are also on the list of ten that made commitments ^[15].

The ICO's other track is formal. "Where organizations expose people to avoidable harm or proceed without adequate safeguards, we will intervene," the regulator said ^[10]. Neither of the two X entities under investigation over Grok is among the ten ^[16].