AWS kills the DIY multicloud VPN as Microsoft bets on AI transparency
7 min read · 10 sources
- AWS Interconnect for Azure is now GA in three regions, replacing DIY multicloud networking.
- Microsoft's AI Transparency Initiative launches with over 40 members to close the agent evidence gap.
- Gartner predicts 70% of production agentic AI will suffer a material incident by 2029 without runtime controls.
- Oracle's Java4Shell patch fixes a flaw demonstrated against Jira, Confluence, Jenkins, and Elasticsearch.
- Intel's expanded Azure deal claims up to 2.3x performance in M-series v4 VMs via CXL.
The day’s biggest news is that AWS finally stopped making you build your own bridge between clouds. AWS Interconnect for Azure is generally available, a purpose-built private link that replaces the DIY multi-layer network stacks engineers have been duct-taping together for years. It lands in us-east-1, ap-southeast-2, and eu-central-1, with Google Cloud and Oracle Cloud Infrastructure support already on the menu.
But the more consequential story is the one about trust. Microsoft gathered 40-plus organizations - Accenture, Adobe, KPMG, Mastercard, Volvo Cars - into an AI Transparency Initiative aimed at closing the “evidence gap” that stops enterprises from independently verifying what an AI agent actually did. That gap is about to get expensive: Gartner predicts that by 2029, 70% of organizations running agentic AI in production will suffer a material incident because of weak runtime controls. The industry is running out of runway to pretend otherwise.
Gartner says 70% of organizations running agentic AI in production will suffer a material incident by 2029 due to weak runtime controls.
AWS Interconnect: the managed bridge between clouds
AWS Interconnect is exactly what it sounds like: a managed service that connects your VPC to an Azure virtual network over private links, no VPN tunnels or third-party transit appliances required. You provision it through the Console, CLI, or API, and AWS handles the resilience and routing between the two clouds.
The interesting part is what this replaces. Teams building multicloud today stitch together VPN gateways, dedicated circuits, and network appliances, then maintain that stack like it’s a second job. Interconnect folds that into a single managed offering, which means fewer moving parts to break and fewer security review tickets. For anyone running workloads split across AWS and Azure for redundancy or data residency, this kills a whole class of “it was working yesterday” networking problems. It’s GA in three regions now, so this is a real option, not a preview promise.
Microsoft's AI Transparency Initiative: closing the evidence gap
The AI Transparency Initiative is Microsoft’s answer to a question every enterprise security team is asking: how do you audit an agent that acts on your network? The answer today is “you mostly can’t,” because models and agents don’t leave the kind of evidence trails that compliance frameworks demand.
The initiative wants a common, open approach to AI security and governance across vendors, not just Microsoft’s stack. That’s a big deal for engineers because it means the audit story for agents might finally converge on something standard rather than 40 proprietary dashboards. The members - Accenture, Adobe, KPMG, Mastercard, Volvo Cars - are the kind of names that make CISOs pay attention. The risk it addresses is concrete: autonomous agents with broad access are an insider-risk nightmare, and the controls to catch them misbehaving are nowhere near mature.
The CIO, CISO, and CRO problem nobody owns
Speaking of blurred lines, a CIO.com piece makes the case that agentic AI is erasing the boundaries between the CIO, CISO, and CRO. When an agent makes a decision that costs money or exposes data, nobody has a clear mandate to own that outcome. The governance models we have assume a human made the call.
The recommended fix is sensible: give agents identities, scoped access, separate control layers, circuit breakers, and human approval for high-impact actions. The hard question the article raises is who that human approver is. Is it the CIO who owns the system, the CISO who owns the risk, or the CRO who owns the financial outcome? Pre-agreed tie-breaks and explicit decision mandates sound like bureaucracy, but they’re cheaper than discovering the answer after an incident. Gartner’s 70% figure is the stick; this is the carrot.
A bottle-cap manufacturer deflates the AI hype
The Register ran a reality check from an unexpected source: a Chinese bottle-cap manufacturer. Their presentation challenged the assumption that every process is ready for AI automation, pointing at aging equipment, limited digital skills, and thin margins. The pitch was that successful adoption starts with understanding employees’ practical knowledge and operational constraints, not with buying a platform.
It’s a useful corrective for anyone who thinks AI is a universal solvent. The factory floor is where the rubber meets the road, and if the data isn’t there or the machines can’t talk to the software, the model doesn’t matter. The lesson for enterprise engineers is the same as for a bottle-cap plant: start with the constraint, not the model.
Gemini Tasks: Google's agent goes autonomous
Google is rolling out task-handoff features in the Gemini app, free for hundreds of millions of users in the US, Japan, Korea, Canada, Australia, and India. Powered by the new Gemini 4 Hydrogen model, the assistant can make phone calls, book appointments, and analyze data - utility bills, say - using integrations like Google Password Manager and Personal Intelligence.
For engineers, the notable bit is the autonomy. The agent acts across services without you driving every step, and it discloses to businesses that it’s an AI and offers opt-out options. That’s a meaningful shift: consumer agents are now doing what enterprise agents are being asked to do, which means the governance problems are moving from the boardroom to the living room. The infrastructure to run these things at scale is the same infrastructure you’re already operating.
Windows endpoint security goes GA without the kernel
At Ignite 2026, Microsoft announced the Windows endpoint security platform is reaching general availability on Windows 11. The pitch: third-party security vendors can protect systems without relying as heavily on kernel-level software. Developers get an API to warn users, limit functionality, or block applications if a PC fails a security check.
This matters because kernel drivers are where security products have historically gone to die - and where they’ve taken down entire fleets with them. Moving protection out of the kernel is a resilience win, and Microsoft frames it as part of the Driver Quality Initiative. The timing also tracks: the window between vulnerability discovery and exploitation is shrinking, so anything that reduces the attack surface and the crash surface at once is worth a look.
Governed agents on Windows: identity, containment, management
Microsoft’s enterprise deployment guide for governed agents on Windows is the practical companion to the transparency initiative. It defines three OS primitives: identity via Microsoft Entra Agent ID, containment via Microsoft Execution Containers (MXC), and manageability via Intune and Entra. MXC enforces platform-level boundaries, from lightweight process isolation up to separate Windows sessions.
This is the architecture that makes the transparency initiative work on the ground. An agent with an identity, confined to a container, and managed by your existing policy engine is an agent you can actually audit. Without those primitives, you’re back to hoping the agent behaves. The guide is 15 minutes of reading that pays for itself the first time an agent does something unexpected.
Java4Shell: the patch you need to test today
Oracle patched a Java vulnerability that lets unauthenticated attackers execute code or crash systems, and researchers have already demonstrated attacks against Jira, Confluence, Jenkins, and Elasticsearch. Exposure depends on the underlying Java version and CPU configuration, so the blast radius is wide but not uniform.
This is a “fix this ASAP” grade bug, which means your patch Tuesday just got busier. The affected applications are the backbone of a lot of enterprises, and an unauthenticated RCE in that stack is the kind of thing that makes for a bad month. Test the patch against your workloads before the weekend, because the exploit code is already out there.
The knowledge layer: domain ontology for agents
Neo4j’s post on domain ontology makes a clean point: “churn” has three different definitions across bank departments - 3.2%, 1.1%, 4.7% - and no agent can tell which applies to a given question. The knowledge exists; it’s just not in a consumable shape. An ontology layer gives agents governed definitions, rules, and provenance so they stop guessing.
This is the boring infrastructure that makes agents actually reliable. The model is only as good as the meaning it’s reasoning over, and if the meaning is ambiguous, the output is a coin flip. For anyone building agent systems across business functions, this is the part that separates demos from production.
Intel and Azure: EMIB packaging and a 2.3x CXL claim
Intel and Microsoft expanded their long-term agreement, with Microsoft as the first customer for Intel Foundry’s EMIB advanced packaging on Azure first-party silicon. The deal also covers multi-year Xeon supply, including next-gen Diamond Rapids CPUs, plus collaboration on CXL, Intel QAT, and confidential computing.
The numbers are worth quoting: up to 2.3x performance in Azure M-series v4 VMs with CXL, 2.5x faster SQL backups with QAT, and memory/register encryption via Intel TDX. Those are the kind of gains that show up in your cloud bill as fewer VMs for the same work. The EMIB angle is strategic: it puts Intel Foundry inside Azure’s own chip designs, which is a bet on the foundry business paying off beyond x86.
You May Also Like
Microsoft's outage week ends with a 15.7% price hike in India
Microsoft spent Thursday through Saturday restoring Exchange Online, Teams, and Copilot after an outage that began with servers failing automated health checks. …
Serval Hits $3B, Oracle Bills Per Task, and GitHub Adds 4M Cores
Serval secured $150 million at a $3 billion valuation to replace legacy IT service desks with autonomous agents, while Oracle announced plans to bill enterprise …
SAP’s Physical AI Hits Factory Floors, Microsoft Authenticator Flaw Forces Immediate Patch
SAP unveiled a Physical AI platform that wires robots, sensors and quadrupeds straight into ERP, turning shop‑floor actions into auditable data. Microsoft …




