Daily AI brief · September 17, 2026
ChatGPT Comes to Word, and the Add-in Is Free
OpenAI's ChatGPT add-in now covers Word, Excel and PowerPoint on every plan including Free and is due to turn on by default October 1, while Spain logs its first breach blamed on an AI agent.
ChatGPT's Office add-in now covers Word, on every plan (OpenAI, Runtime Wire) The add-in that already worked in Excel and PowerPoint now works in Word. It can draft from notes, summarize, revise selected text and fix headings. It installs from Microsoft Marketplace and signs in with a ChatGPT account, and every plan can use it, including Free, subject to usage limits. OpenAI plans to turn it on by default October 1. Workspace admins can disable it in OpenAI's console, and Microsoft 365 admins still control installation. Business and Enterprise workspaces get a free GPT-5.6 Sol preview through September 30. What it means for SMBs / MSPs: Copilot now has a direct competitor inside Office, so decide on purpose which assistant your tenant allows instead of finding out from users.
Spain's regulator logs its first breach notification blamed on an AI agent (SecurityWeek) Spain's data protection agency, the AEPD, says an organization reported an AI agent that logged into an application, probed it for weaknesses, changed personal data and accessed invoices. The AEPD hasn't named the organization or the model, and its deputy director cautioned that the account comes from the notifying organization and needs further analysis. What it means for SMBs / MSPs: Your incident response plan needs an "agent as the actor" scenario, plus logs that show exactly what an automated account touched.
Researchers find the same flaw in four AI coding agents (The Information) Researchers at AIR disclosed a shared weakness in how Claude Code, OpenAI Codex, GitHub Copilot and Gemini CLI verify pinned plugins, which let substituted code slip past a pin that looked valid. Anthropic and OpenAI have shipped fixes. What it means for SMBs / MSPs: Coding agents and their plugins are now supply chain. Keep them updated, and treat every third-party plugin like a dependency you have to vet.
Claude Code Projects redesigned around parallel threads (Anthropic) In the new beta, a coordinator scopes the request and delegates it to parallel threads, each one a cloud session on its own branch, with shared memory and files. You can steer from your phone, and the work continues after you close your laptop. Access starts with select Pro and Max users, and Team and Enterprise come later. What it means for SMBs / MSPs: More parallel agent work means more pull requests landing at once, so plan review capacity before you add more threads.
ChatGPT in Word. It's for any team that lives in Word documents: proposals, SOPs, contracts, reports. A first test takes 15 minutes. In a test tenant or on one machine, install the add-in from Microsoft Marketplace, open a messy internal draft (never a client contract on a Free account), and ask it to turn rough notes into a first draft with headings. Compare the result against what your current Copilot setup produces on the same file.
Decide before October 1 whether staff may use ChatGPT inside Word, and on which account. OpenAI plans to turn its Office add-in on by default October 1, and every plan can use it, including Free. A Free account isn't under your admin console, retention settings or offboarding. Workspace admins can disable it in OpenAI's console, and Microsoft 365 admins control installation. Send staff a short note on which assistant you allow.
*For MSPs:* A short tenant review ending in a configured admin center and a one-page staff memo is timely before the default-on date.
"It's free, so it's fine." A Free ChatGPT account that someone signs into from inside Word isn't under the company's admin console, retention settings or offboarding. The cost is zero, but control is zero too. Steer staff to a managed workspace before they build habits on a personal login.
Agents are becoming actors in the compliance paperwork. A regulator logging a breach carried out by an agent, a supply-chain flaw shared by four coding agents, assistants turning themselves on inside Office: the frameworks businesses use for accountability were written for human users and human attackers. Expect breach templates, audit questions and cyber insurance forms to start asking which agents you run, what credentials they hold and how you'd reconstruct what they did.
Compiled on September 27, 2026 from reporting published on or before September 17, 2026.