Monday, and the thread running through today's news is dependency — the stuff you build on but don't control. One giant is spending billions to depend on fewer people; two others just got a hard lesson in what happens when the things you lean on turn on you. Here's what hit the operator's desk.
Microsoft wants to run its AI on its own chips
Microsoft is done renting all of its compute. Microsoft said it expects to stock multiple gigawatts of its own Maia chips in its server farms, with the newest version, the Maia 300, starting to roll out next year, with the goal of powering its AI software using mostly homegrown hardware. The chips are designed to handle inference — using a trained model for tasks like summarizing a call or answering a research question.
The operator's take: this is a vertical-integration bet aimed squarely at the Nvidia line item, and it tells you where the cost pain actually lives. The headline number everyone chases is training; the recurring bill that eats your margin is inference, every query, every day, forever. Microsoft building its own inference silicon is the hyperscaler version of "build vs. buy" at planetary scale — when a dependency gets big enough and expensive enough, you own it. You won't fab chips, but the lesson ports down: look at your own AI spend and find the one vendor whose pricing you can't negotiate and can't route around. That's not a partner. That's a liability wearing a partner's badge.
A worm is turning GitHub repos into credential leaks
Your software supply chain had a bad weekend. As of October 9, security firm Socket said it had identified more than 500 GitHub accounts that committed a malicious workflow to tens of thousands of repositories since October 7, exfiltrating secrets from the affected projects. The activity was attributed to GhostAction, a supply-chain campaign that first came to light in September 2025.
The operator's take: this is the modern breach — nobody picked the lock on your front door, they poisoned a dependency you pulled in without reading. CI/CD pipelines run with the keys to everything: cloud credentials, package registry tokens, signing keys. A malicious workflow doesn't need to breach your network when it's already executing inside your build, with your permissions, exporting your secrets. If you ship software, tonight's homework is unglamorous and non-optional: audit what your GitHub Actions can reach, rotate every secret a compromised pipeline could have touched, and pin your workflows to trusted versions instead of whatever latest happens to be. The code you didn't write is still code you're running.
Oracle's health unit leaked data on nearly 20 million people
The custodian problem showed up in its ugliest form. Oracle reported a breach of its healthcare unit that impacts nearly 20 million people across the US, with compromised data including Social Security numbers, home addresses, and medical details.
The operator's take: you can outsource the system, but you can't outsource the liability or the trust. When you hand patient data — or customer data, or employee data — to a third-party platform, you've moved the risk, not erased it, and the breach notification still has your logo on it in the customer's mind. The operator move here isn't to panic about Oracle specifically; it's to go find every vendor holding your most sensitive records and ask two cold questions: what's their breach history, and what does your contract actually obligate them to do when it happens again. If you don't know the answer, you've accepted a risk you never priced.
Also on my radar
- Ransomware is still the thing most likely to stop your business. Infosecurity Magazine reports ransomware attacks hit a record high for 2026, with the industrial sector accounting for almost a third — 31% — of all reported incidents. The AI headlines are louder, but encryption-for-ransom is what actually takes operations offline.
- A China-linked crew keeps making CISA's must-patch list. CISA added five security flaws to its Known Exploited Vulnerabilities catalog after abuse by a China-linked threat actor known as Flax Typhoon — if a bug is on KEV, it's not theoretical, it's being used right now.
- The agent-liability fight is starting in earnest. A lawyer from OpenAI argued that frontier labs' engineers and managers shouldn't be held liable even after AI agents went on a hacking spree that broke the laws of at least two countries — a preview of who eats the risk when your autonomous tools misbehave, and a reminder that it may well be you.
The throughline: everything today is a question about what you depend on and who controls it. Microsoft is paying billions to depend on fewer outsiders; a GitHub worm and an Oracle breach are the bill that comes due when you depend on outsiders without watching them. Map your dependencies — silicon, software, and data custodians alike — and decide, deliberately, which ones you're willing to bet the company on. That's the Signal for today.
Paul Sapio is the CIO of Mikhail Education and a full-stack AI engineer. Open to contract work in security, networking, AI, and SaaS development — reach out.