Three Clouds Is Too Few
Build it yourself, let a company do it for you, or rely on open source. People have argued about that for decades, and the answer has always been somewhere on a spectrum. AI-driven attacks moved where I stand on it, and I think putting so much of the internet in three hyperscalers has made it more dangerous and more fragile.
What happened this month
On November 17, attackers published two poisoned versions of Microsoft's agent-framework package to PyPI, a package with roughly 200,000 downloads a month. The stealer ran every time Python started, even if you never imported the package, and it went after Azure, AWS, GitHub, Docker and GitHub Actions credentials (BleepingComputer).
Five days earlier, Outlook, Teams and Copilot went down, with Downdetector logging reports from 74 countries and territories (Al Jazeera). It lasted about a day.
Letting a company do it
For a long time, handing infrastructure to a big provider was the safe answer. They had more engineers, more money and more security people than anyone renting from them, and that's still true.
What changed is the attacker's math. Finding a bug used to take a skilled person a long time. Now a model can grind through code for the price of the tokens, and if I were the attacker I'd point it at the target where one hole opens the most doors. Three hyperscalers is three of those targets, and when AWS, Azure or Google Cloud has a bad day, every company built on them has it too, all at once.
By fragile I mean one failure takes out people who never chose each other. When Teams went down this month, two companies that both run on it couldn't talk to each other, even though neither of them did anything wrong. A business that runs its whole operation on Azure goes down with Azure, and a breach works the same way, one compromised provider and every company on it is exposed at once.
Relying on open source
The PyPI attack wasn't a breach of Microsoft's own infrastructure. It was a supply-chain hit on an open-source package with Microsoft's name on it. I already wrote about the dark side of open source, and I understand the risks I take on by building on it. I do my due diligence, but I also think Microsoft's open tools are the juiciest targets for attacks like this, so I tend to stay away from the big providers' tools and build my own.
Where I do pull other people's code, I pin it to a hash instead of a version label, which I covered in What Your Pipeline Pulls. A label can be moved to point at new bytes and a hash can't. A lockfile with hashes for an earlier agent-framework release would never have installed 1.26.1 or 1.26.2, while anything resolving to the latest version picked them up.
The same AI that finds bugs cheaply can read every open codebase on earth, and a bug found in one install is a bug in every install.
Keep at least one closed layer dedicated to defence against these attacks. When its code isn't public, AI-powered attackers find something they weren't expecting. Hiding the code is only part of it. A layer you own is one you can change as fast as you need to. All of that sits on top of good security underneath, as an extra layer for when you do get targeted by someone who really, really wants into your system. If you are going the open source route, it would be very naive to go all open source, as much as I would like to.
There's an honest problem with my own argument too. A hundred small providers all running the same software share the same CVEs, so spreading out who hosts the code doesn't spread out the bugs in it.
So the real answer is choice. Your choice in the provider(s), and the providers' choice in software. If there was only one real option for a self-hosted Git forge, then one vulnerability in that exposes all 100 small providers that use it. So having a robust open software ecosystem is just as important as having a robust small-provider ecosystem, and I believe both directly feed into each other.
Why I still use Pingora
You could call me a hypocrite here. After a whole section about staying away from the big providers' tools, I build on Pingora, an open-source Rust proxy framework that Cloudflare owns and runs. Everything in this post sits on a spectrum, and Pingora is a spot on it I'm comfortable with.
Before I built on it I did my due diligence. I looked into how it's governed, the community around it and the vulnerabilities it's had in the past, and I was really happy with what I saw. Cloudflare chose Rust so the proxy would be memory safe, and when a request-smuggling bug in Pingora's caching came in through their bug bounty in April 2025, they had it mitigated within 22 hours and published the CVE with a full timeline (Cloudflare).
Compare that with Microsoft this month. Their preliminary report says a bug from a fix they started rolling out in July sat in their gateway servers for four months, then took Outlook, Teams and Copilot down when Teams traffic failed over onto gateways that already had about 14% of their capacity out of service (365fanatic). They also say it was no cyberattack, which I have no way to check. So I'm comfortable building on Pingora, and I keep Microsoft out of my company's stack.
Building it yourself
A self-hosted forge looks cheap to start. Mine began at $6.95 a month, and it didn't last. Running a software company on it every day, with CI builds and mirrors of CVE databases pulled on a schedule, needs a much bigger machine, and the runner belongs on a second one. It's $40 a month now, which is still cheap.
Where I stand now
I think a hundred small providers is a much healthier internet than three hyperscalers. On paper a small provider has fewer resources and fewer staff, and that's true. But everything I said above about juicy targets cuts the other way too. A small provider is far less likely to be picked for a targeted attack. The automated sweeps still hit it like everyone else, because a scanner hits everything running the vulnerable version, and that's where a closed layer earns its keep.
What really bothers me is what the big three do with their resources. Microsoft cut 10,000 jobs in 2023 (Microsoft), more than 15,000 in 2025 (Fortune) and about 4,800 more this July (KSL). Amazon cut 27,000 in 2023 and another 14,000 corporate jobs in October 2025 (PBS). Over the same stretch Satya Nadella's pay hit a record $96.5 million for fiscal 2025 (Fortune) and $90.6 million for 2026 (The Register). In 2024 the US Cyber Safety Review Board reviewed the 2023 break-in to Microsoft's Exchange Online and concluded that "Microsoft's security culture was inadequate" (CSRB report). I think that money could have kept a lot of people employed and bought a much stronger security posture, and I think greed is what led them here.
The stakes are different for a small provider. When a hyperscaler has a massive breach, it's too big to shut down and it can afford the legal battles, so even something really bad is a drop in the bucket. When something really bad happens to a small provider, that's the whole business and often someone's whole life, with legal bills they likely can't afford. So I think small providers and their staff are much more likely to have a security-first culture, and to put the effort back into keeping the internet safe. I still use open source everywhere, I still buy from providers, and I build the layers where being one more copy of everyone else is the risk.
Spreading out works at the scale of one business too. A company that splits its operation across a few providers still has the others to reach its customers and keep working when one of them goes down.
The jurisdiction side of this is its own argument, which I made in Digital Sovereignty and the Nation-State.
Owning the responsibility
Moving off the hyperscalers lowers your risk, but it doesn't make it zero. What it does is move the risk onto you.
When your software ships a security release, nobody patches it unless you do. When your runner pulls a poisoned package, it's your secrets on that box and your job to rotate them. Backups only count if you've restored from one, and if your host folds, the exit plan is whatever you wrote down before it happened. When you go down, there's no status page to refresh and nobody to blame, and you're the one who explains it to the people who depend on you.
You trade a risk you can't see or fix for one that's sized to you and fixable by you. Sovereignty means owning the responsibility too.
Storm Buckets is S3-compatible object storage hosted in Canada, built on Garage.
Have comments? Add me on Mastodon and shoot me a message.