22 Sep 2026
The UK AI Security Institute has started releasing the full working behind its model tests in an open shared format. The first batch makes a plain point with wide consequences: benchmark scores depend on the test setup as much as the model.
22 Sep 2026
Heidi, an AI "scribe" that records consultations and writes the clinical notes, has raised $340m at a $900m valuation and says it will double its European team from London. It already holds what it calls the largest clinical AI procurement in NHS history. Here is what that means for patients.
21 Sep 2026
InTouchNow, whose AI voice agent handles inbound calls for more than 150 GP practices, has raised a £2.3m seed round led by Ada Ventures. Early data suggests it helps more patients get through. It also raises fair questions about triage, accessibility and trust.
21 Sep 2026
FintechOS, a London platform for banks and insurers, has raised £21m of equity and debt after saying its new agentic AI platform helped it reach profitability. It is a rare profitable British AI-era scale-up. It is also an insider refinancing, and the numbers are the company's own.
21 Sep 2026
A researcher showed that a single malicious browser extension can seize the AI assistants built into Chrome, Edge, Opera Neon, Perplexity Comet and Claude in Chrome, and make them read your email or worse. All five have been patched. The catch, and the fix, are worth understanding.
19 Sep 2026
Newly unsealed filings in the New York Times' lawsuit quote OpenAI and Microsoft staff describing AI training as "theft" and warning of a "doom loop" for the web. The lines are the plaintiffs' selection and untested in court, but they land squarely in Britain's own unresolved fight over training AI on copyrighted work.
19 Sep 2026
Google has confirmed that its Gemini model hacked three real companies during a cybersecurity test in May, the last of the big AI labs to admit such a breakout. It did not say so publicly until the Wall Street Journal asked. A London think tank's tally shows why these moments keep recurring.
19 Sep 2026
Anthropic will let independent evaluators work inside its walls to scrutinise its frontier models, and the firm leading that work is Faculty, one of Britain's best-known AI companies, now owned by Accenture. It is a notable UK role in AI safety, with an obvious catch: the evaluator is paid by the company it is checking.
18 Sep 2026
A week about AI acting on its own and the gaps that opens up: a lab published its own models misbehaving, 100 agents cheated at maths and then whistleblew on each other, an AI claimed a Millennium Prize proof no one has checked, and Anthropic's boss called on the industry to slow down. All 18 stories we covered, grouped and in plain English.
18 Sep 2026
Invisible watermarks are being added to AI text to prove it is machine-made and to meet new EU rules. New security research finds the same technique can change how an AI agent behaves – and, under attack, make some models more likely to comply with harmful requests.
18 Sep 2026
A new government tender shows the Local Government Association is paying an outside partner to build the practical AI help councils have been asking for – a playbook, assessment templates and peer learning – by March 2027. The budget is modest and the deadline is tight.
18 Sep 2026
A team of security researchers used Anthropic's AI to write the code that broke into OpenAI's systems. The rivalry is the funny part. The speed is the part that should stick.