AI Weekly for IT Pros: October 24–30, 2026

The most important item for IT teams this week is Microsoft putting the Copilot harness, the agent orchestration layer behind GitHub Copilot, Copilot Studio and the Copilot CLI, into public preview in Microsoft Foundry, with free runtime until December 1. It arrived alongside cheaper models for computer-use tasks and a fresh reminder that coding agents will follow any instruction file they find in their working directory.

AI in Microsoft & Azure

Copilot harness details for platform teams. The harness is selected when you create an agent in Foundry Agent Service and cannot be switched later; existing agents are unchanged. Long-running resilience for hosted agents is now generally available, but your code still has to save progress and avoid duplicate actions on restart.
GPT-6.1 Terra is generally available in Microsoft Foundry. It is available through the Responses API with no SDK changes, with Global Standard deployments in all supported regions and Data Zone Standard in the US, EU and APAC. Data Zone pricing is higher than Global Standard, so pick the deployment type based on your data residency needs.
Azure grows 48% as Microsoft reorganises around agents. Microsoft reported Azure revenue of $33.2 billion for the quarter ending September 30 and now reports an “Agents and Infra” segment. Expect Microsoft’s sales and licensing conversations to lean harder on agent workloads in the months ahead.

Tools for admins and developers

Copilot session continuity and projects. GitHub says the same cloud agent session can now be followed across the GitHub app, VS Code, the CLI and the web, and Copilot projects run parallel agent work across multiple repos. Check which repos and secrets those sessions can reach before turning them on broadly.
Claude API request limit raised to 128 MB. Anthropic raised the maximum request size for the Messages and Token Counting endpoints from 32 MB to 128 MB, which helps pipelines that send large PDFs or images. Context-window limits still apply, so large payloads can still be rejected on token count.
Kimi K3.1 Pro brings a 1M-token open-weight model. Moonshot released Kimi K3.1 Pro with a one-million-token context window and says full weights are due by November 4. Treat it like any open-weight model: test it on your own tasks and review licence terms before production use.
Cohere Rerank 5 for internal search and RAG. The new reranker works on top of keyword, vector or hybrid search without re-indexing, and Cohere says its calibrated relevance scores make cut-off thresholds easier to set. It is a low-effort upgrade if your internal knowledge bot returns noisy results.
1Password for Agents. 1Password launched a way for AI agents to sign in with a user’s saved logins after approval, without the agent seeing the password or one-time code. It is a better pattern than pasting credentials into agent prompts, though launch support is limited to a few agents for now.

AI security and compliance

A rogue AGENTS.md in Downloads can hijack coding agents. Researchers reported to BleepingComputer that agents launched from ~/Downloads auto-load instruction files found there; in one case this led to installing a rogue npm package, and a related CLAUDE.md test exposed an OAuth token. Tell developers to run agents only from project directories and keep approval prompts and sandboxing on.
Warp discloses a breach exposing agent logs and telemetry. The agentic terminal said data was exposed through a third-party vendor rather than its own systems. Agent logs can contain commands, file paths and snippets of code, so check where your developer tools send telemetry and whether you can turn it down.
123 Reg reports an “AI-driven” attack on customer data. The GoDaddy-owned UK registrar says an intruder read customer contact and account data and that it found AI attack tooling on the attackers’ infrastructure; attribution is still a suspicion. If you manage domains there, watch for phishing that uses the leaked details and confirm two-step verification is enabled.

Try this week

Given the AGENTS.md research, look for agent instruction files sitting outside your real project folders. On Windows, this PowerShell checks your home folder and one level of subfolders, including Downloads:

$names = 'AGENTS.md', 'CLAUDE.md', 'GEMINI.md', '.cursorrules'
Get-ChildItem -Path $HOME -Depth 1 -File -Force -ErrorAction SilentlyContinue |
    Where-Object { $names -contains $_.Name } |
    Select-Object FullName, Length, LastWriteTime

On macOS or Linux, the equivalent is:

find ~ -maxdepth 2 -type f \( -name AGENTS.md -o -name CLAUDE.md -o -name GEMINI.md -o -name .cursorrules \) 2>/dev/null

Open anything you did not create yourself before an agent can read it, and make a habit of starting agents from inside the project directory, never from your home folder or Downloads.

Sources

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

Share this content:

AI Weekly for IT Pros: October 24–30, 2026