A quieter Sunday than yesterday, but the throughline is sharper for it: the software you trust to build and run the business is, this week, both the thing under attack and — occasionally — the thing causing the outage. The perimeter you were defending has moved inside your own toolchain.
The attackers climbed the supply chain — to your AI vendor
Most of the AI security conversation is about what models do in the wild. This one is about the labs themselves. In its latest threat report, Anthropic describes criminal groups increasingly targeting AI vendors' own infrastructure — including an attempt to steal a pre-release Claude model. The same reporting notes a Russian threat actor using AI to build, test, and deploy exploits against hundreds of organizations worldwide.
The operator's take: you now have to reason about your AI provider the way you already reason about your cloud and your identity provider — as a high-value target whose breach becomes your breach. Model weights, fine-tuning data, and the prompts your business runs through them are all concentrated in a handful of vendors that are now squarely in the crosshairs. That doesn't mean rip the tools out; it means treat the AI vendor like any other tier-one dependency: know what data leaves your walls, read the security posture, and have a plan for the day the provider — not you — is the one that gets popped.
Microsoft's own September patches are knocking systems offline
Patch Tuesday was last week's wall of work. This week it's the collateral damage. Windows admins report that the September 2026 security updates are causing Remote Desktop Services failures on Windows Server 2019, 2022, and 2025, blocking connections and in some cases requiring a hard reset to recover. Separately, update KB5124008 is knocking some Windows 11 enterprise clients off Always On VPN after the Patch Tuesday rollout.
The operator's take: this is the tax on the "patch everything, now" instinct, and it's why blanket auto-approval on server rings is a bad policy no matter how scary the CVE count. Remote access and VPN are exactly the services your remote workforce and your admins depend on — break them and you've created a self-inflicted outage on top of the threat you were patching against. The discipline is boring and non-negotiable: ring your deployments, hold a canary group for 24–48 hours on anything touching RDS, VPN, or domain controllers, and keep a tested rollback ready. Speed on internet-facing edges, caution on the plumbing your business logs in through.
CISA flags an exploited GitLab bug — your build platform is exposed
The dev platform is production too. CISA added a critical GitLab vulnerability, tracked as CVE-2026-85706, to its Known Exploited Vulnerabilities catalog, meaning it's not theoretical — someone is using it now.
The operator's take: a source-code platform holds the keys to everything downstream — credentials, CI/CD runners, deploy tokens, the pipeline that ships your product. Compromise there isn't one app; it's the ability to poison every app you build. If GitLab is self-managed and reachable from the internet, this belongs at the top of today's patch list, not next sprint's. And if you don't actually need it exposed to the whole world, this is your reminder to put it behind SSO and an allowlist so the next exploited flaw isn't a public-facing one.
Also on my radar
- Hardware wallets, soft targets. Trezor disclosed a phishing campaign that targeted 347,000 email addresses and hooked 2,500 users who clicked a malicious link — the wallet held, but the customer list is the real product being attacked.
- A $320M "white-hat" heist. Alleged white-hat hackers drained $320 million from Liquid's federation wallet, then demanded a bug fix — a reminder that "we did it for the disclosure" is not a defense your incident report will accept.
- Quantum is a planning problem now. The G7 urged organizations to start preparing for quantum cyber threats — harvest-now-decrypt-later means the clock on your long-lived secrets already started.
The pattern across all of it: the trust boundary keeps moving inward. Last week the danger was agents on the wire; this week it's your model vendor, your patch pipeline, and your source-of-truth for code — the very things you'd point to as the foundation. Defense in 2026 means assuming the tools are targets too, and building so that no single dependency's bad day becomes yours. 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.