The AI Leadership Brief

A memory maker, a processor maker and an IT-services firm report growth; SK hynix says supply is not keeping pace and Intel's CEO prioritises more supply

A large cloud provider says 155 responses from one hosted model contained other users' data

In the provider's incident report, a hosted coding model returned 155 responses containing data from other users' sessions between 27 Aug and 9 Sep 2026, all to very long prompts. The model was offered with a larger context window than its cache could hold; the provider says, “We were aware of this mismatch before launch.” A bug in the open-source SGLang engine then let those requests read leftover data from earlier requests. A researcher reported it on 8 Sep; the provider made the model private within about 40 minutes of declaring an incident, found “no evidence of deliberate exploitation”, and says no other hosted model was affected. The provider is not named because the news is negative.

Preprint: where a hidden instruction sits can change whether AI agents obey it

The authors built three agent testbeds (tool use, coding and daily workflows) and moved the same injected instruction between locations. They report that agents trust some parts of their environment more than others, which they call the “trust region effect”. In one result, an agent powered by OpenAI's GPT-5.4 followed an unsafe instruction placed in a repository documentation file (README or TODO) 2.6 times more often on average than the same instruction placed in a code or data file. Not peer reviewed.

MeitY releases a draft handbook for training AI on sensitive data without handing it over

Status: marked “Draft for Discussion”; a framework, not a law. Scope: the handbook on the Data Empowerment and Protection Architecture (DEPA) for AI Training, prepared with the iSPIRT Foundation and steered by MeitY, proposes that training happens where the data sits: in its main design, raw data stays with the training-data provider inside secure data environments and is not exposed to those using it for training. It names roles including training-data providers and consumers and data-exchange platforms. It calls itself a “living resource”, invites “actionable recommendations” and gives no deadline. As reported by ETV Bharat, MeitY Secretary S Krishnan unveiled it at a roundtable.

Runware plans an assembly line for factory-built AI data-centre pods

As reported by TechCrunch, Runware's chief executive says the company plans to scale its Romania factory to 30 MW of data-centre capacity per day by 2028 (“nearly 1 gigawatt per month”), with a second factory due early next year and US expansion to follow. The liquid-cooled pods need power and internet but no water connection. Runware aims to have pods online within 48 hours of delivery (installation and setup, not manufacturing) and cites $1 billion of committed interest covering more than 20 MW. Company plans and claims, not measured output.

Only rounds of $100M or more, announced 26 Oct–1 Nov 2026. Amounts and leads come from each company's announcement; where we could not open it, a bylined report quoting it is linked. Total $1.27B is our sum of the four announced amounts; the “business problem” column is our short description. Counsel note supports only (R16b). No valuations.

Hitachi to launch a 35,000-person digital services company with physical AI as a named focus

Hitachi plans to launch Hitachi Digital Inc. on 1 April 2027, uniting GlobalLogic and Hitachi Digital Services: about 35,000 professionals in 27 countries serving more than 880 active clients. Named focus areas are physical AI, data and intelligence, cloud and site reliability engineering, and enterprise applications. An announced plan; the company is not yet operating.

Caution: verify before business use. This brief summarises the public sources linked beside each item. Forecasts, plans, preliminary results and maker's claims are the sources' own. Business lens lines restate the linked source unless marked “Our summary”; Ask your leadership team lines are the editor's questions, not facts or advice. Headlines, summaries, role tips and leadership actions are the editor's wording. Verify any item, figure or quote independently against the original source before using it for business decisions. Small red reference numbers (R1, R2…) point to the companion Source Register.

Found this useful? Give it a 👍 on LinkedIn and share it with a colleague.

AI model makers are named; the cloud provider in the incident item is not named because the news is negative. Every item links to its source.