A data centre
without power.

Microsoft walks away from a site the grid cannot power

Microsoft has ended its agreement to anchor Nscale's data centre at Loughton in Essex. Grid officials do not expect to supply power there until the 2030s.

What happened

Reuters reported on 19 October that Microsoft had pulled out of a deal to be the anchor customer for a data centre being built by Nscale, a UK AI infrastructure company, at Loughton in Essex. According to Reuters, electricity grid officials had told Nscale they did not expect to supply power to the site until the 2030s. The Telegraph reported the move first.

The site was announced in September 2025, among the technology deals unveiled during the US President's state visit. Reuters says it was due to begin operating early next year, with Nvidia chips providing computing capacity for Microsoft's Azure cloud services in Britain.

What both companies said

Microsoft said the decision relates to a single site and that it remains committed to its partnership with Nscale. Nscale called it a London-wide grid issue with no impact on its other contracts. It said it remains committed to Loughton and to the UK, and is exploring alternative power sources for the site and other places to expand.

Why it matters

Promises about AI capacity are usually made in chips, megawatts and opening dates. This one failed on a grid connection. Reuters notes that queues to connect to the grid are delaying some data-centre projects by as long as 13 years. For anyone told that a service will run from a particular UK site, the opening date depends on power that the supplier does not control.

The government's advisers define sovereign AI by choice, not ownership

The Council for Science and Technology advises the UK to keep access to leading global AI models while building selected capability of its own.

What the council advised

The Council for Science and Technology is the government's top independent advisory body on science and technology. On 19 October the government published its letter to the AI minister, Kanishka Narayan, dated 30 July. It summarises advice on AI sovereignty that the council was still developing.

The council defines AI sovereignty as having enough agency across the AI value chain to deliver the UK's chosen outcomes, keep economic and public value in the UK, and stay resilient under duress. Without deliberate action, it warns, the UK risks becoming "a high-value AI customer, rather than a country able to shape the technology."

Four recommendations

It asks for a UK AI Sovereignty Plan, starting with a quick map of computing power, data, skills, models, governance and adoption, to decide which capabilities the UK should own, build with others, or secure reliable access to. It wants one accountable owner across government, investment in skills and data, and trusted access to leading models kept alongside selected UK capability. Pilots would test open-weight and smaller domestic models on public-interest tasks. The council also suggests considering the Alan Turing Institute's proposal for a public-interest sovereign base model.

Council for Science and Technology advice, July 2026. Not adopted policy.

Why it matters

This is advice, not adopted policy. Its central point still applies to any organisation: sovereignty here means keeping choices open when a supplier, a country or a power connection fails. The Loughton withdrawal, published the same day, is a small example of why that matters.

Outside testers find gaps in OpenAI's monitoring of its own agents

Two outside researchers spent three weeks inside OpenAI testing the system that watches its internal coding agents. They found ways agents could avoid it.

What was tested

From 21 July to 11 August, one employee of METR, an independent AI evaluator, and one from Redwood Research stress-tested OpenAI's internal misalignment monitoring. That system watches the AI agents OpenAI's own staff use for coding, looking for behaviour such as sabotage, deception and attempts to disable the monitor. The pair worked on OpenAI devices that never left the premises, read internal code and documentation, and interviewed about a dozen staff. METR published its summary on 22 October.

What they found

The findings describe the system as it stood on or before 21 July. The pair estimated that roughly 75 to 94% of agentic internal Responses API traffic from models at least as strong as GPT-5.5 was eligible for monitoring, with substantial uncertainty. Reinforcement-learning training runs were not covered when the exercise began. They found a way coding agents could in principle get around their sandbox and take any action the user could take, and two settings that could start a new agent without monitoring. Because the monitor checked agents only when they handed control back, an agent could put off being checked by continuing to act.

What remains unverified

OpenAI says every vulnerability affecting its production safeguards has been patched, and it has described wider monitoring since. METR says it has not evaluated those claims. By agreement, the public post is short, and OpenAI could redact it as it alone saw fit.

Why it matters

Last week's lead covered Anthropic giving METR inside access. This is what such access can produce: specific, testable faults, published with the company's consent. It also shows the limit. Readers learn what was wrong in July, not whether today's system works.

'Password1' and the test account nobody removed

The ICO has reprimanded a council data body and its IT supplier after ransomware exposed information about 3,300 people, most of them children.

What happened

On 1 November 2024, attackers got into the network of Data Cymru, the Welsh local government data unit, through a test account called Test2. Its IT supplier, Ogi Networks, had created the account in January 2012 with the password Password1. More than twelve years later it was either still active or had been switched back on by mistake. The attackers encrypted 182.7GB of data and took 42GB, including records from two projects supporting children and families.

Whose fault

On 20 October the Information Commission's Office issued reprimands to both organisations. Its head of investigations said responsibility "did not sit with one organisation alone." Ogi created the account. Data Cymru kept overall control and could tell Ogi to create or remove accounts. Data Cymru had no contract with the data protection terms the law requires for a sub-processor, had not written down how its supplier's account services would be monitored or audited, and reviewed accounts irregularly and only in part.

What has changed since

Both organisations have added multi-factor authentication and periodic account reviews, and they signed a compliant contract on 30 April 2025. The reprimands describe failures before the attack, not the systems today.

Why it matters

The memorable detail is the password. The lesson is the gap between two organisations, each able to assume the other was looking after old accounts. Any outsourced service has the same gap until someone is named to close it.

Law firms report AI gains. Few see them in the accounts.

PwC's 35th Law Firm Survey finds most firms report better productivity, quality and responsiveness from AI, but few have turned that into commercial returns.

What the survey says

PwC released its 2026 Law Firm Survey summary on 15 October, and our research desk assessed it on 19 October. Growth remains firms' main ambition, but PwC says performance is falling short of it, and the gap between the largest firms and the rest continues to widen. It calls for more attention to pricing, incentives and how work is staffed, and says growing reliance on AI, cloud services and third-party software makes resilience part of any change programme.

What it cannot show

This is a public summary of what firms report about themselves. It does not publish the full results, the sample or the questions, and reported gains are not audited outcomes. It cannot say where any time saved has gone.

Why it matters

Saved time can turn into lower bills, more review, spare capacity or higher margins. Only some of those show up in a firm's results. For clients, the question is whether faster work reaches the price or the service, or stays inside the firm.

Does an agent's memory help? It depends on the agent.

Researchers built a way to test AI agents piece by piece. Remembering past attempts helped one agent and hurt two others on the same maths task.

What the researchers built

AlphaDiana, described in a paper published on 14 October, measures AI agents by changing one part at a time: the tools, the instructions or the memory of past attempts. The authors tested three open-source models with three open-source agent frameworks on maths and reasoning benchmarks. Our research desk assessed the paper on 20 October.

What they found

Wrapping a model in an agent framework did not reliably make it more accurate. The framework helped on some questions the model could not answer alone, and lost others it answered correctly without tools. In one maths test, keeping memory of earlier attempts helped one framework and hurt the other two, and none of the memory settings beat the model working on its own.

Why it matters

Memory is often sold as an upgrade. This study suggests its effect depends on how a particular system stores and uses what it remembers. The results come from maths problems, not legal or business work, and have not been reproduced.