Three stories worth your coffee this Friday. China's biggest AI lab got kicked out of federal agencies and raised a pile of money two days later. A Chinese hacking crew turned an open-source AI agent framework loose on a supply chain. And Microsoft says its diagnosis bot beats doctors, with some fine print you should read.
01 of 03
DHS Boots DeepSeek, DeepSeek Raises $15 Billion Anyway
On October 20 the Department of Homeland Security published an order in the Federal Register (document 2026-21447) telling federal civilian agencies to stop buying DeepSeek products and pull the ones they already have out of government IT systems. It also covers third-party products built on DeepSeek tech. Agencies can ask for exceptions. DHS's reasoning is that Chinese law could make data from U.S. government users accessible to Chinese authorities.
This stretches an existing rule. A December 2025 law already kept DeepSeek out of intelligence and national security systems. Now the rest of the civilian government is covered too. Beijing's Foreign Ministry spokesperson Mao Ning said China opposes "overstretching the concept of national security and politicizing technology and trade issues."
Two days later, Bloomberg reported DeepSeek closed roughly $15 billion, about 100 billion yuan, with battery giant CATL and Tencent among the investors. That comes after a June round of more than $7 billion at a $50 billion valuation. No valuation has been reported for this one.
Here is where I land. The ban makes sense on its own terms. I would not want federal employees pasting agency data into a chatbot that answers to Beijing either. But if anybody in Washington thought this would slow DeepSeek down, the market answered in 48 hours. Chinese money does not care whether a federal agency buys a license. Keeping it off government laptops is security hygiene. It is not a strategy for competing with them.
02 of 03
Chinese Hackers Ran AI Agents Through ConnectWise
Microsoft published a detailed write-up on October 21 about Storm-0940, a China state-aligned group active since 2021. The group breached ConnectWise's cloud backup service and stole the application credentials that connect ConnectWise to its customers' Microsoft 365 environments. With those keys, the attackers could walk in as a trusted app without anybody typing a password.
The tool doing the heavy lifting was ARTEX, an open-source penetration testing framework that can plug into any AI model. It is a multi-agent system that runs attacks from reconnaissance all the way to data theft with little human input. Microsoft says Storm-0940 started using it in September.
The numbers are what got my attention. Between October 10 and 14, one ARTEX-driven password spray hit 327 accounts at eight organizations with more than 151,000 failed logins, peaking around 2,900 failures a minute on October 11. Then on October 16, the same day ConnectWise posted its advisory, the group used the stolen app credentials to try signing into more than 2,250 customer tenants. Most attempts failed. A few landed privileged cloud access, some within three hours of first contact. In one case a ConnectWise app assigned itself Global Administrator. Microsoft is careful to say a successful sign-in does not by itself prove data was stolen. ConnectWise has paused its cloud backup access while it investigates.
What I want to know is how many small IT shops are still running with those app connections live. That is the real lesson here. The attack itself is old school: steal keys, log in, poke around. The AI part is the speed. When a bot can try a few thousand doors a minute, all day and all night, "we'll look at the alerts Monday" stops being a plan. If you are an MSP on ConnectWise, go read Microsoft's checklist today, not next week.
03 of 03
Microsoft Says Its AI Beats Doctors, With Asterisks
Microsoft AI published a study in Science on Thursday about MAI-DxO, a system that strings several AI agents together to act like a panel of doctors. They question a simulated patient, order tests, and argue it out before landing on a diagnosis. On hard cases pulled from the New England Journal of Medicine, DxO got it right 83% of the time. Primary care doctors and general internists, working alone with reference materials, got 36%.
Dig into the results and the picture gets more honest. They tested models from OpenAI, Google, Anthropic, Meta, and DeepSeek, and the best runs came from OpenAI's GPT-5 and GPT-5.5. GPT-5 by itself already scored 78.9% on the complex cases. Wrapping it in DxO added about four points and cut roughly $3,000 off the cost per case. GPT-5.5 went from 72% to nearly 87%. Google's Gemini 2.5 Flash jumped from 46.4% to almost 69%, but still trailed the stronger models. On emergency room cases, DxO barely moved GPT-5 at all.
Mustafa Suleyman, who runs Microsoft AI, is calling the goal "medical superintelligence" and says health models will get "an order of magnitude better in capability" within 24 to 36 months. Next up is a real-world test with Beth Israel Deaconess Medical Center in Boston in emergency and urgent care, with consenting patients. Microsoft is also building a healthcare foundation model with Mayo Clinic, due next year.
I will give them this. The study is in Science, not a blog post, and they published the parts that make DxO look less magic. But the headline compares AI to a generalist doctor flying solo on cases that specialists and teams normally handle. That is not who you see when you are really sick. Most of the gain came from the underlying model, not Microsoft's wrapper. Show me the Beth Israel numbers on real patients and I will get excited. Until then, it is a very good test score.