LodeHQ
AI news for people who shipFree, by email

DevOps · Wednesday, 26 August 2026 · Issue 84 · 8 stories

Hobbyist kills idle servers, BMCs hide attackers

Two stories that make you rethink what's running when you aren't looking: Hobbyist puts your home projects to sleep, and BMC flaws let attackers live beneath the OS. Also, Terraform 1.16 imports inside modules.

Containers & Orchestration

Hobbyist puts your self-hosted projects to sleep, wakes them on demand

Postgres and apps on a home server burn power and memory even when nobody touches them. Hobbyist is a v0-alpha, no-K8s orchestration layer that runs each project in containers, shuts them down when idle, and cold-starts them on demand through a Postgres-wire-protocol and HTTP proxy, roughly 170ms for a database wake on the author's hardware. One command makes a project; Caddy handles TLS and Tailscale covers access. The broken bits are listed in the README, and there's no hosted tier planned.

github.com ↗14 min read at source

Observability & Reliability

Atlassian's RCA bot replaces the on-call correlation grind

At scale, paging someone to eyeball dashboards, logs, and traces and manually guess the culprit eats the most expensive minutes of every incident. Atlassian treats the whole problem as a correlation across three axes: signal type, time, and the service dependency graph. Anomaly detectors plug in modularly, and the engine emits ranked hypotheses so responders go straight to validation. The blast-radius-first approach narrows hundreds of services to tens before analysis even starts.

cncf.io ↗10 min read at source

Cloud & Platform Engineering

Your AI agent shipped to prod, nobody told it not to. LocalStack makes the case for a sandbox

An AI coding agent will happily provision a bucket, wire a Lambda, hammer a production table, and call real AWS on the way out. The bill lands two days later. LocalStack argues the missing layer is an infrastructure sandbox, and breaks down four kinds you can't confuse: code execution, model, data, and cloud. The one most teams skip is the one that actually stops the damage.

blog.localstack.cloud ↗12 min read at source

Terraform 1.16 lets you import inside modules and run natively on IBM Z

The big one: `import` blocks now work inside modules, so teams can fold existing infrastructure in without a separate top-level step. It's also the first release shipped as a native s390x (zLinux) binary, unlocking mainframe-class compute for Terraform runs. Around the edges, new `store` blocks carry ephemeral values across plan and apply, and lifecycle gets `destroy = false` to lock a resource down. Action triggers also gain `on_failure` modes instead of a hard halt.

github.com ↗3 min read at source

DevSecOps

Two guest-to-host escapes in one month, patched at fleet scale with zero customer impact

DigitalOcean shipped its first Januscape livepatch 45 minutes after the page, closing thousands of hypervisors within 48 hours on kernels where the upstream fix even existed. Then a second, harder AMD issue hit the 1,600 hosts that couldn't be livepatched, forcing a full kernel update and reboot three weeks later. The lesson worth stealing is the long tail: a few hundred 5.10-kernel hosts had no upstream patch, so the team wrote new livepatches from scratch. A credible account of what real fleet-scale security operations look like under pressure.

digitalocean.com ↗14 min read at source

BMC Holes Let Attackers Live Beneath the OS, Past Every Endpoint Tool

Your servers' Baseboard Management Controllers run outside the OS, and that's exactly the problem. Researchers warn that flaws in BMC firmware and the aging IPMI protocol let attackers take control below the operating system, surviving reinstallations and ordinary recovery. Standard EDR and antivirus sit above this layer, so a compromised BMC is invisible to the very tools meant to catch it. For GPU clouds and bare-metal fleets with thousands of hosts on shared management networks, one foothold can reach far beyond a single box.

infoq.com ↗3 min read at source

Apache Maka keeps your AI agent's work on your machine, sandboxed

Maka, now incubating at Apache, is a local-first agent workspace that runs models and tools behind a sandbox boundary while leaving sessions, tool calls, and recovery data on your hardware by default. Tools that poke outside the sandbox need your approval, and every turn is written to a durable record you can branch, search, and resume from. One Runtime Host drives the desktop app, a CLI/TUI, and an eval harness, so your data stays yours whichever interface you pick.

github.com ↗7 min read at source

A scanner found, patched, and Google shipped: agents-cli 1.4.1 closes a credential-exfil hole

An automated scanner didn't stop at flagging a bug in google/agents-cli, it wrote the fix. The arbitrary file write in scaffold create could smuggle credentials out of an agent workspace, and Google merged the patch into v1.4.1. No CVE, no advisory, just a scanner closing the loop from finding to shipped patch. That is the scanner role evolving from reporter to fixer.

github.com ↗1 min read at source

You’re caught up.

That was all 8 stories in the 26 August issue.