AI Daily Brief: 25 June 2026

25 June 2026

Today's brief has a clear thread: AI is becoming infrastructure, not just software. The most useful stories are about control, cost, privacy and operational resilience, because those are the issues now deciding whether AI creates business value or quietly adds risk.

OpenAI and Broadcom reveal Jalapeno inference chip

OpenAI and Broadcom have announced Jalapeno, OpenAI's first inference-focused ASIC, with engineering samples already running in the lab at target frequency and power. OpenAI says the chip was co-developed from initial design to manufacturing tape-out in nine months, with its own models used to accelerate parts of the design and optimisation process.

The important business signal is not the chip name. It is OpenAI's intent to control more of the stack: chip architecture, kernels, memory systems, networking, scheduling, deployment systems and product experience. For UK firms, that points to a market where model choice, cloud choice and hardware economics become more tightly linked.

Our take: This is another sign that frontier AI is turning into vertically integrated infrastructure. Buyers should expect better performance over time, but they should also watch for deeper vendor lock-in and less transparent cost structures.

AI coding agent bills are becoming a board-level cost issue

Gartner has warned that consumption-based pricing for AI coding agents is creating sharply variable costs for software teams. Its analyst Nitish Tyagi said bills that once looked like $20 or $100 can jump to $2,000 to $5,000 per developer per month, with extreme cases reaching $20,000 in token charges.

Gartner also argues there is no direct relationship between higher token consumption and higher productivity. The recommendation is straightforward: use context engineering, route simpler work to smaller models, reserve frontier models for high-value tasks, and measure AI coding spend like any other material production cost.

Our take: The free trial era is over. Engineering leaders need usage controls, routing policies and cost-per-outcome reporting before AI coding agents become another uncontrolled SaaS line item.

Infrastructure teams report AI-caused incidents at scale

A Spacelift-backed survey of 406 IT decision makers found that 93% of organisations had experienced AI-caused infrastructure incidents, while only 19% had the governance needed to respond properly. Reported consequences included reworking AI-generated changes, security misconfigurations reaching production, compliance violations, infrastructure drift and incidents caused by agentic systems.

The uncomfortable detail is the governance gap. The same research says 86% of respondents believe they can govern AI, but only 30% have a formal AI governance policy in place. That gap is where production risk lives.

Our take: AI-generated infrastructure code should not bypass ordinary change control. The practical answer is governed infrastructure-as-code, automated validation and metrics that track AI-generated changes before they hit production.

Microsoft uses AI analysis in malware supply-chain takedown

Microsoft, security partners and international law enforcement disrupted infrastructure behind StealC and Amadey, taking down or blocking more than 200 domains and command-and-control servers. Microsoft says AI analysis helped investigators connect the two malware operations and support a wider RICO-style civil case.

The broader Europol-linked campaign also flagged and restricted more than $47 million in cryptocurrency assets and recovered about 27 million stolen credentials. Microsoft said Amadey and StealC were linked to more than 140,000 infected computers globally in just the first two weeks of May.

Our take: This is a useful example of AI improving defensive investigation speed. For businesses, the lesson is that attackers are industrialised, and defenders need better data correlation across identity, endpoint, network and legal response.

Medical AI privacy research exposes patient-level risk

German researchers reported that some medical diagnosis AI models can be probed to infer whether a patient's data was used in training. The research looked at seven medical AI datasets across images, ECG records and electronic health records, and found near-perfect attack success for some individual patients under membership inference attacks.

The risk is sharper for underrepresented groups, because outlier records can be easier to identify. For healthcare providers and vendors, this raises the bar for privacy audits, differential privacy and consent processes when patient data is used to train or validate models.

Our take: Anonymised data is not automatically safe data. Any organisation using health, HR or sensitive customer data for AI training needs individual-level privacy testing, not just aggregate risk reports.

Met Police expands live facial recognition into London's West End

The Metropolitan Police will start using static live facial recognition cameras in London's West End and Soho by the end of 2026, following a six-month pilot in Croydon. During that pilot, the Met said more than 470,000 people walked past the cameras, resulting in 173 arrests and one false alert.

Civil liberties group Big Brother Watch criticised the expansion, arguing that permanent biometric surveillance of public space is incompatible with policing by consent. The debate matters beyond policing because it is another test of whether the UK can deploy high-impact AI systems with enough public legitimacy.

Our take: Public-sector AI cannot rely on operational benefits alone. Accuracy, watchlist governance, appeal routes and parliamentary oversight will decide whether these deployments earn trust or trigger backlash.

Google Search users get new AI data training opt-out controls

WIRED reported on Google's new AI data training controls for Search, explaining how users can opt out of having certain Search interaction data used to improve AI features. The move lands as AI search products continue to blur the line between ordinary search usage, assistant behaviour and training feedback.

For businesses, this is another reminder that user settings, retention policies and data use disclosures are becoming part of AI product design. If customers cannot understand what feeds model improvement, trust erodes quickly.

Our take: AI data controls are moving from privacy-page detail to front-line product expectation. Companies building AI features should treat opt-outs, audit trails and plain-English explanations as product requirements.

Quick Hits

• VentureBeat reported that Mistral launched OCR 4, positioning document extraction as a full enterprise AI workflow rather than a narrow conversion tool.
• VentureBeat also covered Mindstone's model-routing memory work with Rebel, aimed at helping enterprise agents remember which model is best for each task.
• The Register reported that Microsoft rivals are submitting evidence to the UK watchdog on where Microsoft's software ecosystem has hurt competition.
• WIRED reported on continued backlash to A24's Google AI collaboration, showing how creative AI partnerships remain commercially sensitive.

Frequently Asked Questions

How often is the AI Daily Brief published?

Every morning at 7:30am UK time, covering the previous 24 hours of AI news from over 30 sources.

How are stories selected?

UK-relevant stories are prioritised first, then by business impact and practical implications for UK organisations adopting AI.

Why should business leaders follow AI news?

AI is moving faster than any technology in history. Staying informed is essential for making smart decisions about AI investment, adoption, and governance.