[go: up one dir, main page]

All about clouds, AI and that devops

Kubernetes pod hardening: a container is still a Linux process

Start with a quick check on one of your clusters: 1 2 kubectl run probe --rm -it --restart=Never --image=busybox:1.37 -- \ grep -E 'Seccomp:|CapEff' /proc/self/status Unless someone has already changed the defaults, you’ll see something like this: 1 2 CapEff: 00000000a80425fb Seccomp: 0 Seccomp: 0 means no syscall filter at all. The process can ask the kernel for anything its capabilities allow, and that kernel is shared with every other pod on the node....

November 5, 2026 · 10 min · Oleksandr Kulbida

Building Docker images with Nix: same inputs, same hash

Run docker build twice on the same Dockerfile, with no changes in between, and compare the results: 1 2 docker build -t app:v1 . && docker save app:v1 | sha256sum docker build -t app:v2 . && docker save app:v2 | sha256sum The two hashes won’t match. Nothing in your code changed, but the image did. I’ve been building images with Nix for a while, and it fixes exactly this....

November 1, 2026 · 12 min · Oleksandr Kulbida

Anthropic replaced git with its own source control: Scion

Anthropic doesn’t use git for its monorepo anymore. In five months they moved to their own version control system called Scion, and the reason is simple: agents write too much code for git to keep up. The numbers In May 2026 about 2,500 changes a day were merged into Anthropic’s monorepo. By September it was over 30,000. Most of that growth comes from Claude writing code, not from people. At that scale the repository itself becomes the problem:...

October 19, 2026 · 6 min · Oleksandr Kulbida

Five layers of LLM defense, and why activation guardrails won't save you alone

If you think LLM attacks are still caught with a list of regular expressions, you’re living in 2023. A lot has changed since then. Defenses now live everywhere from the training run to the production proxy in front of the model, and some of them reach right into the model’s hidden states while it’s thinking. I find it useful to split all of this into five layers, from what’s baked into the weights to what happens around a deployed model:...

October 19, 2026 · 13 min · Oleksandr Kulbida

AX: Google's orchestrator for AI agents on Kubernetes

In 2014 Google released Kubernetes to manage lots of containers across lots of hosts. Today it’s the de facto standard for running workloads. In 2026 Google released AX (Agent Executor) to manage lots of autonomous agents inside a Kubernetes cluster. Whether it becomes a standard too, only time will tell. AX is infrastructure for running AI workloads. It runs on top of Kubernetes and Agent Substrate, which provides the sandboxed execution....

October 9, 2026 · 6 min · Oleksandr Kulbida

k8s-aibom: finding out which AI workloads are actually running in your cluster

At some point someone from security or compliance asks the platform team a simple question: “Which models are we running in production, and where?” Usually nobody has a good answer. There’s a spreadsheet that was correct six months ago, a few Helm charts that mention vLLM, and that one team that deployed Ollama in a dev namespace “just to try something” and forgot about it. k8s-aibom is an open source controller from Google Cloud that tries to answer this question by looking at the cluster itself....

October 5, 2026 · 9 min · Oleksandr Kulbida

50 ChatGPT "commands" for learning faster from books, articles and lectures

Most people use ChatGPT for reading the same way. Paste an article, type “summarize this”, get back a wall of text that’s about 70% as long as the original. Then you don’t read that either. A while ago I came across a list of short “commands” that change what the answer looks like. Instead of “summarize”, you write /flashcards or /argumentmap or /actionplan, and the output comes back in a format you can actually use....

October 3, 2026 · 10 min · Oleksandr Kulbida

Docker Sandbox Kits: a Dockerfile for what your agent is allowed to touch

Docker published the Sandbox Kit Specification, an open format for packaging an environment for an AI agent together with a list of what that environment needs from the outside world. A Dockerfile tells you what’s inside the image. A Kit also says what the image will want from outside: network destinations, credentials, volumes, tools, and context for the agent. The Kit doesn’t grant any of it, though. It asks. The runtime decides what to give and enforces the limits....

October 2, 2026 · 8 min · Oleksandr Kulbida

systemd-report: a new systemd subsystem and lower-level alternative to node_exporter

systemd has a new tool called systemd-report. It gathers facts and metrics about a machine, writes them into one timestamped JSON document, optionally signs it, and can upload it to a server you choose. This is for people running fleets of servers. If you have one laptop, you’ll probably never touch it. What goes into a report There are two kinds of data. The first is stuff that barely changes, which systemd calls “facts”:...

September 29, 2026 · 4 min · Oleksandr Kulbida
auth-svc in the old default namespace forwarding to auth-svc in the new authentication namespace

Moving a Kubernetes service to another namespace without downtime

CNCF published a nice write-up by George Sims: Migrating a critical Kubernetes deployment from the default namespace without any downtime. The story is about an auth service that lived in default for years, but the approach works for moving any service from one namespace to another. default is just the example. The main idea is simple. You keep the old DNS name alive as a compatibility point. In the old namespace you leave a Service of type ExternalName that points to the service in its new namespace, and callers keep working without changing anything....

September 25, 2026 · 6 min · Oleksandr Kulbida