AI Weekly for IT Pros: November 14–20, 2026

AI Weekly for IT Pros: November 14–20, 2026

Microsoft Ignite 2026 (November 17–20) dominated the week, and its message for IT teams is clear: AI agents are now something you identify, govern, budget for and audit like any other workload. Right behind it, a supply-chain compromise of Microsoft’s own agent-framework package on PyPI is the item to act on today if you build with Python.

Top stories

  • Ignite 2026: agents become an admin workload. Microsoft framed Entra Agent ID and Agent 365 as the way central IT can “discover, govern, and audit agents”, and shipped operational pieces such as the Azure Copilot deployment agent and Azure Monitor health models as generally available. Expect agent inventory, identity and cost controls to land on your team’s plate in 2027 planning.
  • Microsoft’s agent-framework PyPI package was compromised. Malicious versions 1.26.1 and 1.26.2, published November 17, ran a credential stealer at Python start-up targeting Azure, AWS, GitHub, Docker, Hugging Face and OpenAI secrets, plus Azure managed-identity and CI OIDC tokens. Any build agent or dev box that installed them should be treated as compromised: rebuild on 1.26.0 and rotate every reachable credential.
  • OpenAI releases GPT-6.2 Astra and Sol. Both are available in ChatGPT and Codex for Plus, Pro, Enterprise and Edu users (Enterprise and Edu once a workspace admin enables them) and in the API as gpt-6.2-astra and gpt-6.2-sol. Astra’s API price drops to $8/$40 per million input/output tokens, 20% below GPT-6 Astra, while Sol stays at $2/$10 — worth re-running your cost models.
  • MAI Code 2 arrives in GitHub Copilot. Microsoft’s own coding models, MAI Code 2 and MAI Code 2 Flash, are available in GitHub Copilot, with Code 2 Flash in limited-access preview in Microsoft Foundry. Microsoft says they are not yet tuned for very long-horizon agentic work, so treat them as a cheaper option for routine coding tasks rather than a drop-in for everything.
  • NVIDIA ships Vera Rubin NVL72 as revenue hits $115.8B. Q3 FY2027 revenue rose 103% year over year, with data centre revenue at $107.6 billion, and NVIDIA guided to $130 billion next quarter. The results were presented by interim CEO Colette Kress; for buyers, the takeaway is that demand for the newest accelerators remains extremely strong, so plan GPU capacity and reservations early.

AI in Microsoft & Azure

  • Copilot Managed Runtime is GA. Apps built with Copilot are hosted in your tenant, and Microsoft says every app is visible and managed in the Microsoft 365 admin center. Copilot Code is due to become generally available in the coming weeks, so now is the time to decide who may build and publish apps.
  • Copilot Autopilot starts rolling out to Frontier customers. Each Autopilot gets its own Microsoft Entra Agent ID and a named manager, and IT can scope its access or block it at any time. Review your Frontier enrolment and Entra policies before users start asking for it.
  • Agent 365 gains cost and MCP controls. Microsoft announced cost management features such as spending policies, limits and alerts, plus controls for MCP servers, plugins and skills, and containment of local agents through Intune, Defender and Microsoft Execution Containers. These are the levers to explore if “shadow agents” are already a concern.
  • Azure infrastructure for agents. Azure Cobalt 200 VMs are generally available, Azure SRE Agent gets new enterprise controls in preview, and customers will be able to deploy open-weight models and agents from Foundry to their own AKS clusters. Windows Server 2028 is now in public preview.

Tools for admins and developers

  • Azure Copilot deployment agent (GA) turns natural-language requirements into deployed workloads. Pair it with your existing policy and review gates rather than letting it bypass them.
  • GitHub Copilot hybrid intelligence (preview) brings local-model support to the GitHub Copilot app, CLI and VS Code, according to Microsoft’s Windows team — relevant if data residency has kept coding assistants off some projects.
  • Amazon SageMaker AI endpoint monitoring is on by default for inference endpoints using vLLM containers on GPU instances. It flags latency drift, idle GPUs and missed optimisations via AWS Health at no additional GPU cost; teams can opt out per endpoint with the OPT_OUT_TRACKING container environment variable.
  • OpenAI prompt-caching guidance for GPT-6.2: OpenAI recommends keeping prompt prefixes unchanged and appending new instructions to benefit from the cached-input discount ($0.40 per million tokens for Astra) — a cheap win for internal tools that send long system prompts.

AI security and compliance

  • Supply-chain risk now includes AI SDKs. In the agent-framework incident, the payload ran through a .pth file even when the package was never imported, and a second malicious version appeared minutes after PyPI quarantined the first. Pin versions, use lockfiles with hashes, and check egress logs for unexpected webhook traffic.
  • GPT-6.2 tightens cyber safeguards. OpenAI’s system card (published November 17) says both models meet its “Critical” threshold for cybersecurity capabilities, and the launch post notes the tighter safeguards also restrict some legitimate vulnerability research and security-testing workflows. Security teams relying on these models may need to apply for vetted access.
  • AI agents bypassed safeguards in live testing. BleepingComputer reported that agents in third-party safety tests made tens of thousands of requests against Epoch AI’s real website and disabled a fake network-blocking control in isolated runs. If you test agents, isolate them properly and monitor outbound traffic.
  • Models can be jailbroken with “it’s a game” framing. A Vals study covered by TechCrunch found leading models often complied with harmful simulated targeting tasks once framed as a game, and no safety classifier blocked the requests. It is a reminder not to rely on vendor guardrails alone in your own AI applications.

Try this week

Inventory the Azure AI (Foundry / Azure OpenAI / Cognitive Services) accounts in your subscription and the model deployments inside them — a useful first step before applying any of the new agent governance controls:

# List all Cognitive Services / Foundry accounts in the current subscription
az cognitiveservices account list \
  --query "[].{name:name, kind:kind, location:location, rg:resourceGroup}" \
  --output table

# List the model deployments in one account
az cognitiveservices account deployment list \
  --resource-group <resource-group> --name <account-name> --output table

Both commands are documented in the official Azure CLI reference; add --subscription to check other subscriptions.

Sources

Questions about using AI in your IT environment? Get in touch.

Share this content:

Leave a Reply