A Week of Outages, and Nobody Attacked Anything
Microsoft 365, the FAA, AWS, Chase, United and the NSA all had outages within six days. In the same week the FBI warned about ARTEX, an AI framework that reaches privileged access in under three hours.
Microsoft 365 was down for about a day last week. Nobody hacked it. A defect in a fix that Microsoft began rolling out to its network gateway servers in July sat there until Teams traffic was moved onto the gateway Exchange Online also uses, and then Outlook went dark worldwide.
Over the next five days the FAA, AWS, Chase, United and the NSA joined the list. None of those has been attributed to an attack either.
What broke
Here's the week as far as the public record goes. NBC News has a good overview.
- Thursday 12 November. Microsoft 365 goes down. Microsoft calls it "a technical issue not related to cyber or AI / agents". Its preliminary report blames the July gateway fix, and the final one is due by 20 November.
- Friday 13 November. The FAA orders a nationwide ground stop that lasts a little over four hours. The agency points at a newly installed router and says it still doesn't have the root cause.
- Sunday 15 November. AWS has authentication and connectivity problems in us-east-1. Its incident update gives no definitive cause.
- Monday 16 November. Chase customers have trouble for most of the working day. The bank blames a software bug in its on-premise hardware systems. United announces a "widespread system outage" the same afternoon and gives no cause.
- Tuesday 17 November. NSANet, the NSA's classified network, has problems for several hours. Apple's iCloud sign-in fails the same day after what Apple calls an internal update.
US officials have been consistent that the incidents aren't connected. National Cyber Director Sean Cairncross told Semafor this was "not any sort of coordinated, nefarious cyberattack".
Why in the same week?
The theory most people in the industry lean towards is rushed updates. AI tooling now finds vulnerabilities faster than anyone can fix them. Microsoft's October Patch Tuesday covered about 1,413 CVEs, the first month ever above a thousand, according to Dark Reading. November's had 1,094 by SOCRadar's count, and other trackers land a few either side. Everyone is patching flat out to stay ahead of attackers, and changes made in a hurry break things.
I'd call that unproven. Microsoft's own explanation is a fix from July that only failed under an unusual traffic pattern four months later, which doesn't sound like a patch pushed out last week. The FAA blames hardware. Some of these may turn out to be ordinary bad luck that happened to cluster.
One of them does fit, though. Asked about the NSA outage, Director of National Intelligence Jay Clayton tied it to patching after a PHP vulnerability was identified. Another official described "temporary degradation during scheduled routine maintenance". Cairncross, in the same interview where he ruled out an attack, said the disruptions were coming from updating and modernising legacy systems.
So the best-supported version of the theory is narrower than the one going around. At least one of the week's outages came from patching, by the government's own account, and the volume of patching has roughly tripled since the summer. I wrote in June that the comfortable patch window is gone. This is what the other side of that looks like. If you can't afford to wait, you test less, and you find out in production.
The FBI's warning about ARTEX
On the Friday of that same week the FBI published a FLASH alert about ARTEX. ARTEX is an open-source framework, written in Go, that hands the whole intrusion to AI agents. They do the reconnaissance, find the weaknesses, exploit them, escalate privileges and take the data out. Its developer published it on GitHub in July as a tool for authorised testing and pulled the repository on 8 October. Forks, archived copies and a commercial edition are still around.
From the alert, as reported by BleepingComputer:
- Some intrusions went from initial targeting to privileged access in under three hours.
- Attacks on South Korean financial institutions exposed personal information on about 67,000 people.
- Attackers compromised the application identities of ConnectWise's cloud backup service, then used ARTEX to try authenticating to more than 2,250 downstream Microsoft 365 tenants with permissions those customers had already granted. Huntress has counted at least 162 tenants actually compromised.
That last one is worth reading twice. The attackers didn't need 2,250 passwords. They needed one vendor whose app had been approved in 2,250 places.
The model that doesn't say no
ARTEX has no model of its own. You plug one in, and the supported list includes Claude, GPT, DeepSeek, GLM, Grok and Kimi. According to The Next Web's reading of the alert, DeepSeek's Flash models are the ones most used on attackers' servers. They're cheap, and they show no visible guardrails when the request is framed as penetration testing.
There's a joke going around about asking one model after another until one of them agrees to help. It's funny, and it's also an accurate description of the situation. When Claude or ChatGPT refuses a request like that, a company chose that behaviour. Nothing in the technology enforces it. If one model says no you ask the next, and anyone can download an open-weight model and run it on their own hardware with nobody watching.
Nothing here is new
What stays with me is that none of the reporting on ARTEX mentions a zero-day. The FBI says the framework is good at finding weaknesses in overlooked systems, and it lists partner portals, employee applications, legacy services and third-party integrations. Amazon's researchers watched similar agents succeed against default and hardcoded credentials, keys left in code and unauthenticated interfaces.
In other words:
- the supplier portal everyone forgot about
- the admin password sitting in a repository
- the app integration somebody approved in 2021 that nobody remembers
We've been writing those findings into audit reports for fifteen years. They used to take a human attacker days of patient work to dig out, which gave you a chance of noticing. Now it takes an afternoon.
Three hours against seventy-two
Under GDPR you have 72 hours to report a personal data breach to your data protection authority. In Iceland that's Persónuvernd. The clock starts when you become aware of the breach.
The attacker needs three hours. Your 72 don't begin until someone notices. So the question I'd put to any organisation is who would notice if somebody got in at three o'clock tonight. If the honest answer is "whoever reads the logs on Monday", the attacker has had the weekend.
The FBI's recommendations are unglamorous and correct:
- List every internet-facing system, including the partner portals and employee apps.
- Keep cloud administration interfaces off the open internet.
- Review which applications have been granted permissions in your tenant, and remove the ones you can't explain.
- Rate-limit sensitive APIs.
- Alert on bursts of automated requests, unexpected credential or privilege changes and large data transfers.
Two risks, one supplier
The week showed me two things. We're now updating fast enough to knock our own systems over. And the people who'd like to knock them over have automated the work.
The ConnectWise case and the Microsoft outage are the same lesson from two directions. If everything you run sits with one supplier, both risks live in the same place. Their bad change takes you down, and their compromised identity lets someone in.
So here's a question for your next team meeting. If your biggest supplier disappears tomorrow morning, what stops working first? And can anyone in the room say how long you'd last without it?