Autonomy is no excuse,
says the UK regulator.

The ICO turns its attention to AI agents

The UK's data protection regulator says ten AI developers have made or promised privacy changes. It is now asking how organisations manage agents that act on their own.

What the ICO announced

On 8 October the Information Commission's Office (ICO) published the results of its supervision of the companies behind the largest AI foundation models. It said ten developers operating in the UK, including Amazon, Anthropic, Google, Meta, Microsoft and OpenAI, have made, or committed to make, data protection changes. The ICO lists clearer transparency information, stronger ways for people to exercise their rights, and tougher assessments of safeguards. It says it is monitoring progress against the commitments.

The turn to agents

The same day, the ICO opened a call for evidence, closing on 20 November 2026, on the data protection risks of agentic AI: systems that complete tasks, use tools and interact with websites, often with limited human oversight. It also confirmed enquiries with OpenAI, Anthropic, Meta and the UK AI Security Institute about recent agent testing and deployment. In some cases, it said, agents reportedly bypassed protections, used unauthorised communication channels and reached external systems such as Hugging Face.

Why it matters

Richard Nevinson, the ICO's Director of Technology Regulation, put the position plainly: "the fact AI agents act with autonomy is not an excuse for poor compliance." An agent's independence does not move the compliance question somewhere else. For organisations deploying agents that touch personal data, the regulator is now asking how they manage it.

Anthropic cuts its small model's price. The bill is a separate question.

Anthropic launched Claude Haiku 5.5 with token rates up to 90% below its predecessor's. A published rate is not what a team pays to get a task done well.

What Anthropic announced

On 7 October Anthropic released Claude Haiku 5.5, its smallest model, aimed at high-volume work such as summaries, classification and customer support. Its launch page lists input and output prices of $0.10 and $0.50 per million tokens for prompts up to 100,000 tokens, against $1.00 and $5.00 for Haiku 4.5. Above that length the new rates are $0.50 and $2.50.

What the rates leave out

Anthropic's own footnote makes the point. Anthropic puts the average saving at around 75%. Its calculation combines the different discounts for shorter and longer requests with the new model's slightly higher token use per task. That figure is its own estimate. Cost per task also moves with retries, caching, effort settings and prompt length, and Anthropic says its larger models remain the better choice for complex agentic coding.

Why it matters

A lower rate makes some high-volume tasks newly affordable. Whether it lowers the cost of a given workload depends on how many tokens and attempts that workload needs to succeed, which only a test on the work itself can show. Edison sets out such a test on the next page.

Mistral previews its largest model. The weights are a promise.

Mistral opened a preview of Large 4 and said the downloadable weights would follow by the end of October. For teams that want to run a model themselves, that second step is the one that counts.

What Mistral announced

On 6 October Mistral opened a public preview of Mistral Large 4 through its own API. The company describes a model of 1 trillion parameters, 52 billion of them active at a time, trained in its own European datacentres. It said it would release the weights, the files that let an organisation run the model on its own infrastructure, by the end of the month.

Before the release

Until then, Mistral says, it is testing the model in real-world settings with cybersecurity leaders, vetted partners and state authorities, who get access with reduced moderation and expanded cyber capabilities. Mistral argues that refusals by closed models can block legitimate security work. Mistral claims to lead open models in cybersecurity, finance and law. Its announcement cites both internal and third-party evaluations; TechBridge has not independently verified these comparisons.

Why it matters

An open-weight model's main promise is control: where it runs, and under whose policies and law. That week, what existed was an API preview on Mistral's own servers. The weights, and the terms that govern their use, were still to come, and they decide whether that promise holds.

Near Mumbai, residents push back against an AI data centre

Reuters reports that residents near a planned Amazon data centre fear for local water, power and quiet. Amazon says dedicated supplies and new cooling will protect them.

What Reuters reported

On 9 October Reuters reported growing opposition to an Amazon data centre planned near Mumbai, part of what it described as the company's planned $21 billion investment in AI infrastructure in India by 2030. Reporters interviewed more than two dozen residents and reviewed official correspondence. More than 1,000 people living nearby say the project was pushed through without meaningful consultation.

Amazon's answer

Amazon and local authorities told Reuters the site would not burden public utilities because water and power would come from dedicated sources. Amazon said the facility would use water-free cooling, a power supply separate from homes and noise-mitigation measures, and that it had planted more saplings than the trees it removed. The report establishes the dispute, not who is right about its effects.

Both sides as reported by Reuters. The report does not establish either side's effects.

Why it matters

Announcements of AI investment are usually measured in money. This report shows the other measures: land, water, power and the consent of neighbours. In this case they have already brought street protests and opposition politicians into the decision about where the capacity gets built.