AI News Roundup — October 20, 2026

Today’s edition is about AI agents inside the organisation — how to control them, what they were trained on, and what they are doing to the people who work alongside them. Microsoft set out how it wants companies to police agents, a group of Canadian publishers took Anthropic to court, and Anthropic published unusually candid data about its own workforce.

Microsoft says AI agents should be treated as insider risks, and is building controls to match. In a post on the Official Microsoft Blog on October 19, Microsoft Security executive vice president Hayete Gallot argued that agents running inside a company need continuous oversight, and the company extended its Project Perception security system to cover what it calls the “agentic insider”. A companion post on the Microsoft Security blog introduces Dynamic Permissions, which give agents task-scoped, time-limited authority rather than copying an employee’s access; Dynamic Agent Risk, a continuously updated risk score; and Dynamic Conditional Access, which can tighten or stop an agent as its risk rises. Microsoft says discovery now covers 60 types of local agents, including Claude Code, OpenAI Codex and GitHub Copilot. Security Affairs notes that Microsoft’s proposed Guardian Agents, which investigate risk using human-written playbooks, could themselves fail or be manipulated. Microsoft’s posts do not say which of these features are generally available. The approach follows the joint NIST, NSA and CISA agent-incident guidance we covered yesterday.

Canadian news publishers are suing Anthropic in Ontario over how Claude was trained. A coalition including Postmedia, Torstar, The Globe and Mail and The Canadian Press filed a claim on October 6 that became public on October 19, alleging that Anthropic infringed their copyright by using their journalism to develop Claude, breached their websites’ terms of service and circumvented measures meant to stop copying, The Globe and Mail reports. MobileSyrup adds that the claim alleges ongoing scraping in response to user prompts, and that the same publishers, with CBC, sued OpenAI in 2024 in a case that has not been settled. Anthropic had not responded to requests for comment, and none of the allegations has been tested in court.

Anthropic says its technical workforce has become markedly less junior. In a research report published October 16, Anthropic says junior staff have fallen from 21% to 14% of its technical departments since early 2025, and from 27% to 9% of its technical hires. Technical teams now make up 56% of the workforce, down from 61%, and the median employee spends 20% more time in meetings than a year earlier. TechCrunch reports that Anthropic attributes the smaller technical share partly to non-technical staff building software with coding tools, and that the company acknowledges more code does not necessarily mean better code.

Our take. Microsoft’s framing is the right one: an agent is a new kind of identity on your network, and it should start with no standing access, get only what a task needs, and lose it when the task ends. That is achievable outside any single vendor’s stack, but only if you first know which agents are running on your endpoints — the kind of read-only inventory Odyssey Sentinel is built to provide. Our guide to running AI agents safely in production covers scoping, approvals and logging in more depth, and if you are designing custom agents or the cloud permissions around them, our consulting services can help — get in touch.

Sources: Official Microsoft Blog, Microsoft Security Blog, Security Affairs, The Globe and Mail, MobileSyrup, Anthropic, TechCrunch