Government bundles, bank rollouts and agent attacks push AI from access to control

October 2, 2026

Microsoft made Microsoft 365 G7 purchasable for US Government Community Cloud customers, bundling Copilot-powered AI and Agent 365 with productivity and compliance tools inside a government cloud offer. That matters less as another Microsoft package than as a sign that AI buying is being pulled into governed suites: the model is no longer the whole product, because access, compliance and daily work tools are being sold together. For customers beyond government, this likely sharpens the same requirement - if AI is to be used in regulated work, it must arrive inside an administrative surface that already controls where data goes and who may use it.

Anthropic and Barclays expanded their collaboration. Claude Code is Anthropic's coding assistant for software work, meaning the tool developers use while writing and changing code rather than a chatbot kept apart from production systems. We read this as another step from trial use to operational dependency: once a bank puts one model family into developer workflows, legacy-modernisation work and customer-service tools at once, model choice starts to shape internal process, training and oversight beyond the vendor that announced it.

The Dutch Institute for Vulnerability Disclosure said attackers using AI agents chained two zero-day flaws in its Zammad helpdesk system, stole contact data and urged users to upgrade or take systems offline. A zero-day is a flaw that defenders have had no patching window for, in other words a bug being exploited before normal defence cycles can catch up. The effect is likely to reach well past one compromised helpdesk: as agents become part of real attacks, organisations need stronger limits around what connected systems can do automatically, and vendors selling AI into work have less room to treat security failure as a separate problem from product design.

OpenAI said Albertsons is expanding ChatGPT Enterprise and OpenAI API use across retail operations, and launched a Safeway shopping experience inside ChatGPT. AWS said Kodiak AI chose it as primary cloud for AI development, simulation and safety validation for autonomous trucking. Proofpoint said TA419 ran credential-phishing campaigns spoofing an Anthropic employee and a former White House official to target AI policy experts. Researchers at the Weizmann Institute unveiled a model that reconstructs viewed images from high-resolution fMRI scans.