Daily AI brief · September 10, 2026
OpenAI opens the Codex agent engine as an API
OpenAI put the harness behind Codex into a public-beta Agents API, while Anthropic's threat report and a Microsoft invoice-fraud warning showed what attackers are doing with the same tools.
OpenAI launches the Agents API in public beta (OpenAI) The Agents API gives developers the harness and infrastructure that run Codex. You define the task, model, tools and compute, and OpenAI handles long-running sessions, orchestration, context compaction, recovery and subagents. Agents can run in OpenAI-hosted sandboxes, on your own infrastructure or with partners including Cloudflare, DigitalOcean, Vercel, Modal and E2B. The public beta offers US data residency but doesn't support Zero Data Retention. What it means for SMBs / MSPs: Building a durable back-office agent no longer means writing your own orchestration layer, but check the data-retention terms before you point it at client records.
Anthropic's threat report: seven Chinese labs, about 190M exchanges (Anthropic) Anthropic's fourth threat intelligence report covers Claude misuse from December 2025 through August 2026. The categories are cyber operations, influence campaigns, surveillance, scams, biological and weapons misuse, and illicit distillation. Anthropic says Alibaba's campaign alone produced more than 151 million exchanges through over 3,500 fraudulent accounts between May and July. What it means for SMBs / MSPs: The report documents scams and cyber operations that smaller actors can now run at scale, so it's a useful source for your security-awareness training.
Microsoft: AI-assisted fake-CEO invoice scam hit over a million inboxes (Microsoft Security Blog) Between August 3 and 5, attackers sent more than a million emails impersonating CEOs of the target companies. Each asked accounts payable for an ACH payment of nearly $50,000, backed by a fabricated invoice and a fake forwarded thread with ServiceNow. About 88% of targets were in the US, and the templates showed signs of AI-assisted construction. What it means for SMBs / MSPs: A polished email with a believable back-story no longer suggests it's real. Payment changes need a callback, every time.
ChatGPT Library adds SharePoint, Dropbox and Box (OpenAI) Paid ChatGPT users on the web can now browse and @mention files from SharePoint, Dropbox and Box without uploading them. Files stay in the source system and citations link back to them. Mobile support comes later. What it means for SMBs / MSPs: If your client's staff use ChatGPT, it can now reach SharePoint directly. Review who has connector rights and what their SharePoint permissions actually expose.
Gemini gets a native Windows app (Google) Google released a Gemini desktop app for Windows 10 and 11 that opens with Alt+Space over any program. It connects to Gmail and Drive, and some agent and media features require a Google AI subscription. What it means for SMBs / MSPs: Google Workspace shops on Windows now get a one-keystroke assistant, and Microsoft 365 shops will see it installed on unmanaged machines.
ChatGPT Library with SharePoint. This one is for M365 shops that already pay for ChatGPT Business or Enterprise. Connect a single test SharePoint site that holds a policy handbook, then ask ChatGPT three questions staff ask every week. Check whether the citations go back to the right file. Plan on 30 minutes, and use a test account with narrow permissions so you see what a normal user would see.
Set a rule that any new or changed bank details get a callback to a known number. Between August 3 and 5, attackers sent over a million fake-CEO emails asking accounts payable for ACH payments of nearly $50,000, backed by fabricated invoices and fake forwarded threads. A polished email proves nothing. Take five minutes with whoever pays the bills, and add a second sign-off above an amount you choose.
*For MSPs:* A 45-minute finance-team session can set up the callback rule, dual sign-off and a mail-flow rule that tags external mail using executives' display names.
Moving regulated data into a public-beta agent service. The Agents API is promising, but it's in beta and doesn't offer Zero Data Retention. Keep health, legal and payment data out of it until the terms fit your client's obligations.
The agent harness is becoming a commodity. OpenAI's Agents API joins Anthropic's Managed Agents and Salesforce's newly previewed Trusted Enterprise AI Harness. Sessions, sandboxes, memory and tool routing are now something you rent. The advantage is shifting to whoever knows the client's process well enough to define the task, the tools and the approval points.
Compiled on September 27, 2026 from reporting published on or before September 10, 2026.