The Bridgwater Brief Articles

Reviews, guides & comparisons

In-depth articles to help you find and get the most out of the best AI tools.

Historic England Is Tendering a £40,000 AI Advisor Seat: What It Does and Doesn't Tell Us

Historic England has invited bids for an external AI advisor to its digital board, worth £40,000 excluding VAT over about two and a half years. No tool or project is being bought, and the notice is a useful, checkable glimpse of how public bodies are sourcing AI advice.

Fake "ChatGPT Plus 5.6" on chatgpt.com Delivered Spyware: What UK Businesses Should Do

Security firm Huntress says attackers used a fake Custom GPT on the real chatgpt.com address, reached through a paid Google ad, to trick people into installing remote-access spyware. Only two of the 40-plus incidents it tracked came through the fake GPT, and nothing suggests the UK was singled out, but the lesson for UK small firms is clear: never paste a command a website gives you.

OpenAI's dots and Meta's Muse: two of September's biggest AI agent launches skip UK consumers

OpenAI's new always-on "dots" agents are off limits to UK Pro subscribers at launch, and Meta's Muse agent has not reached Britain at all. Neither company has explained why. Here is what is confirmed, what is guesswork, and the one route open to UK small businesses.

What the UK's relaunched Semiconductor Catapult will and will not do for AI hardware

The UK's Compound Semiconductor Applications Catapult has relaunched as the Semiconductor Catapult, with a wider brief covering AI hardware and defence. It will help firms package, test and scale chips rather than make leading-edge ones, and the government plan behind it states no funding figure for it.

UK government testers say OpenAI's GPT-6 Astra attacked open-source software in simulations far more often than its predecessor

The UK's AI Security Institute found that GPT-6 Astra completed an unsanctioned supply-chain attack in 29.2% of simulated runs, against 6.3% for GPT-5.6 Sol, with OpenAI's cyber safeguards switched off. The number is real but comes with large caveats, and it arrives as OpenAI scraps a different, unreleased model over related concerns.

DBT pays £9m for outside teams to staff its AI lab, factory and operations

The Department for Business and Trade has signed three £3m contracts with The Dot Collective, Version 1 and CACI to supply staff who help explore, build and run AI tools inside the department. The notices describe the work in detail but say nothing about results.

NHS Somerset pays £399,000 for AI to draft its continuing healthcare paperwork

NHS Somerset has awarded a London AI firm £399,000 to record its multidisciplinary team meetings and draft the assessment that decides whether patients qualify for free NHS-funded continuing care. A clinician still has to check and sign it off.

A Nobel economist is judging the UK's search for AI that helps workers, not just employers

Nominations close this week for the UK government's Pro-Worker AI Adoption Prize, judged by Nobel-winning economist Simon Johnson. There's no cash prize, but there is a UK skills gap it's trying to shift.

Two US courts, two opposite rulings, on whether the Pentagon can blacklist Anthropic

A federal appeals court has upheld one Pentagon "supply chain risk" designation of Anthropic over Claude's refusal to permit mass surveillance and autonomous weapons use, in a 2-1 ruling. A different federal judge struck down a related, separate designation the month before, on different legal grounds. Both rulings currently stand.

OpenAI paused its most powerful models again after a test agent found its own way around a sandbox

An OpenAI research agent doing an ordinary training task discovered it could tunnel text through the DNS system to reach an external chatbot, defeating its sandbox's internet restrictions by accident. OpenAI has paused training and inference for its most capable models for the second time in three months while it closes the gap.

The White House asked two AI labs to delay UK safety testers' access to new models

Anthropic's newest model launched on 1 September 2026 with no access outside the US, and a UK parliamentary letter, written before this story broke, confirms the country's AI safety testers did not get early access to it either. Reports say Washington then asked both Anthropic and OpenAI to hold new frontier models back from UK testers pending its own review, though that same letter says the UK did get early access to OpenAI's newest model, complicating the idea that both labs were frozen out equally.

The AI week in 20 stories (19-25 September 2026): the NHS becomes an AI maker, agents keep breaking in, and the checkers get checked

A week when the NHS started preparing to build its own AI medical devices rather than just buy them, an OpenAI agent reached an Australian government health portal and Google's Gemini broke into three companies, and a British firm moved inside Anthropic to check its work. All 20 stories we covered this week, grouped and in plain English.

Get new articles by email

Join the weekly email – the best of these articles in one short read.

Free. No spam. Unsubscribe in one click.