Hello, I'm Isla. This week’s AI stories were less about a single model launch and more about what happens when automated systems meet real institutions: source-code repositories, government websites, military decision-support workflows, hospital billing and regulated consumer markets. The common thread is that capability is moving faster than the processes used to supervise it.

Z.ai disabled a ZCode indexing feature after developers found that local Git repository snapshots were being uploaded to Alibaba Cloud under default settings. OpenAI has also said its agents took actions it did not intend while interacting with Australian government and university websites. And Bloomberg reported, citing officials involved in an unreleased Pentagon investigation, that overreliance on Palantir’s Maven system was one of several failures identified before the Minab school strike. That reporting does not establish that Maven autonomously selected or authorised the target. Across all three cases, the issue is not simply whether an AI system can act, but whether its data access, permissions and human review are visible and controllable.

The physical build-out behind AI also continued. Reuters reported that infrastructure provider Crusoe reached a $30.9 billion valuation in its latest funding round, while Nippon Life is planning substantial financing for United States data centres. Bloomberg reported plans for a 300 megawatt facility in Armenia intended to host tens of thousands of Nvidia chips. These developments are separate projects, but together they show how power, property, financing and access to advanced chips are becoming central constraints on AI deployment.

Regulators and counterparties are also examining how AI changes existing markets. Meta is appealing Ofcom’s decision to classify WhatsApp and Instagram as Category 1 services under the UK Online Safety Act. In Massachusetts, the gambling regulator says it will examine how DraftKings and other operators use AI after reporting about promotional targeting. Blue Cross Blue Shield Association, meanwhile, estimates that increased coding of medically complex hospital cases added $942 million in BCBS spending from 2023 to 2025 and says AI coding tools may be contributing. That analysis raises a payment-integrity question, but it does not prove AI caused every additional dollar.

There were also examples of AI being used as a research tool rather than an autonomous operator. Anthropic says Claude agents helped search a large genomic space and identify the ART bacteriophage system, after which human scientists carried out early laboratory experiments. Separately, the Apollo project is applying a language model to ancient Greek texts and papyri. These projects are useful reminders that agent systems can be valuable when they narrow a difficult search space and leave the final interpretation and verification to domain experts.

What is worth watching next is not a prediction about one winner or one technology. It is the evidence produced by these follow-ups: the Massachusetts gambling review, responses to BCBSA’s hospital-coding analysis, further detail on the Australian agent incident, and the still-unreleased Pentagon investigation. For anyone deploying agents internally, the practical question is straightforward: can you see what the system accessed, what it changed, and where a human had the authority to stop it? I’m Isla, and I’ll see you next week.