Monday’s AI news is about the plumbing under the agents. Microsoft says its own Maia chips will become the main engine behind Copilot and its other AI products, and a third generation is coming in 2027. Britain announced a package to harden its defenses against AI risks, including plans for incidents involving autonomous agents. And Salesforce is buying Loops, a startup whose AI watches your business numbers and tells you why they moved. The through-line: agents are getting cheaper to run, harder to ignore, and closer to the tools you already use.
1. Microsoft’s Maia 300 is built for agents, and Maia will run most of its AI
In a post on the Official Microsoft Blog, Scott Guthrie said Microsoft will scale its Azure Maia chip family to multi-gigawatt deployments starting in 2027, making Maia the primary inference system for Microsoft products and services.
The current chip, Maia 200, is already in production for Microsoft Copilot. GitHub Copilot and Microsoft Foundry are set to follow by the end of the year. Maia 200 runs OpenAI models, Microsoft’s own MAI models, and open-source model families, and it’s deploying across the U.S., Europe, and Australia. Microsoft says it delivers up to 1.7x more tokens per watt for built-in AI in Teams, Copilot, and Cowork compared with leading GPU infrastructure.
Maia 300 is the third generation. Microsoft calls it a “complete intelligent system” for long-context reasoning, latency-sensitive conversations, and sustained, multi-step agentic tasks. It uses an inference-first design, and Microsoft expects 40% more tokens per megawatt than leading GPU-based systems, plus up to 7x the rack-scale token throughput of Maia 200 in a full 96-accelerator setup. Production deployments begin in 2027. More details are promised at Microsoft Ignite 2026 next month.
Why it matters: you’re not buying this chip. But if you pay for Copilot, GitHub Copilot, or build on Foundry, the cost of running long, multi-step agent work is the thing that decides pricing, limits, and speed. More tokens per watt is how agents stop being a luxury feature.
Sources: Official Microsoft Blog (Oct. 12, 2026); Microsoft Source EMEA.
2. Britain steps up its defenses against AI risks, including rogue agents
The UK government announced a package of measures today aimed at protecting the country from increasingly powerful AI. It builds on two pilots that used frontier AI to help find and fix over 1,000 potential vulnerabilities across government systems, and that program is being extended.
Speaking at Base One, Imperial College London’s science and tech incubator, AI Minister Kanishka Narayan called for a whole-society approach to resilience. “I want Britain to become one of the hardest countries in the world to attack using artificial intelligence,” he said. His message to frontier labs was blunt: listen when your scientists warn of serious dangers, investigate unexpected behaviour, and “if you cannot adequately control the risks, do not release the system.”
What’s on the table:
- Exploring stronger protections under the Cyber Security and Resilience Bill so essential services like hospitals and utilities are better protected, including against AI loss-of-control risk.
- Work with the National Cyber Security Centre and the AI Security Institute on how to respond to incidents involving AI agents, which the government warns could play out faster and wider than other major incidents.
- Figuring out how near misses get identified and reported.
- AI Security Institute evaluations of the latest advanced models, after work with leading developers and the US Government.
Cyber Minister Liz Lloyd said it’s now “business-critical” for organisations large and small to be on the front foot on cyber security.
Why it matters: governments are starting to plan for agent incidents as their own category, not just another hack. If you run a small organisation, the government’s own notes point to the basics: Cyber Essentials and signing up for the NCSC’s Early Warning Service.
Sources: GOV.UK (Oct. 12, 2026); Minister Narayan’s speech.
3. Salesforce is buying Loops to give Tableau an agent that explains your numbers
Salesforce signed a definitive agreement to acquire Loops, an AI-powered analytics platform that keeps an eye on business performance, explains why key metrics change, and surfaces ways to improve using machine learning. The plan is to fold it into Tableau as part of Salesforce’s agentic analytics push.
The interesting bit is how it shows up. Salesforce says Loops has “headless, agentic capabilities,” so it can be triggered right from Slack and email to monitor KPIs, catch trends and anomalies, and flag what’s driving a change.
“Loops accelerates this mission, using powerful agentic AI to uncover the drivers behind business performance,” said Mark Recher, EVP and GM of Tableau. Loops CEO and co-founder Tom Laufer said the idea is that “analytics should help businesses understand what drives their results and decide what to do next.” The deal is expected to close in early November, subject to customary conditions. Salesforce didn’t disclose a price.
Why it matters: “why did this number move?” is the question every ops lead and owner asks on Monday morning. An agent that answers it inside Slack is a lot more useful than another chart nobody opens.
Source: Salesforce (Oct. 12, 2026).
4. What to watch: who pays when personal agents go wrong
A Quartz analysis updated today argues agents are finally having their breakout moment, with Meta’s Muse topping 2.5 million U.S. downloads in two weeks and OpenAI’s Dots arriving at DevDay. It also flags the open questions: a reviewer who clicked “Allow Always” had Muse send a Marketplace buyer his home address, Sam Altman expects “multi-level” AI liability like the auto industry, and Cloudflare’s CEO has floated charging bots per page load.
Why it matters: permissions, liability, and who covers the bandwidth will shape which personal agents you can actually trust.
Source: Quartz.
Quick take for builders and small businesses
- Watch your AI bills, not chip specs. Efficiency gains like Maia’s tend to show up as better limits and pricing on tools you already pay for.
- Put an agent on your KPIs. Have AI watch a few key numbers and explain moves in plain English.
- Treat agent incidents as real. If a government is planning for them, your business should at least know what its agents can touch and how to shut them off.
- Read the permission prompt. “Allow Always” means always. Keep money, addresses, and replies to strangers behind your approval.
- Get the basics done. In the UK, Cyber Essentials and the NCSC Early Warning Service are a cheap first step. Everywhere else: patch, back up, use MFA.
Cheaper agents, tougher rules, smarter dashboards. The week’s off to a busy start.
Share this article
Leave a Reply