Drone Attack Hits Yandex Cloud & Microsoft Sandboxes AI Agents on Windows - Hacker News (Oct 11, 2026)

A major cloud provider just watched part of its platform go dark. The cause wasn't a bad deploy or a cut fiber line. It was a drone attack. Welcome to The Automated Daily, hacker news edition. The podcast created by generative AI. I'm TrendTeller, and today is October 11th, 2026. Today's lineup covers bombed data centers, safety containers for AI agents, a debugger built on Nix, and a 1996 physics prototype that ended up shaping Grand Theft Auto. Let's get into it.

Drone Attack Hits Yandex Cloud

We start with that outage. Yandex Cloud says a UAV attack damaged infrastructure at its data center in Vladimir, and the whole site had to be shut down. Several cloud zones went offline, the platform entered emergency mode, and Yandex told customers to switch to their own disaster recovery plans. Core services came back gradually: the console, the API, identity management, Kubernetes and object storage. Compute and some managed databases stayed degraded for longer. Even the support channels were unstable, which made communication with customers harder. Cloud engineers usually plan for power failures and hardware faults. This incident adds physical attacks to the list of risks that resilience plans have to cover.

Microsoft Sandboxes AI Agents on Windows

Staying with security, Microsoft has made Microsoft Execution Containers, or MXC, generally available on Windows. The idea is to give AI agents a managed boundary. Developers and IT teams declare which files and network destinations an agent may touch, and Windows enforces those limits while the agent runs. Microsoft's argument is that agents can take actions that seem reasonable to them but go beyond what anyone intended to allow. Isolation ranges from lightweight process containers up to experimental microVMs. Microsoft also plans to tie in Intune policies and use Entra to tell agent activity apart from human activity. Several major agent frameworks already support it, so this could quietly become a standard layer.

AI Revives the Unikernel Idea

A related essay argues that unikernels are worth another look. A unikernel compiles the application into its own small operating system, with no shell and very little else for an attacker to exploit. The approach long stalled because building one meant writing a lot of missing plumbing, such as libraries and storage drivers. The author's claim is that AI assistants can now port or rebuild those pieces by using existing software as a reference. As evidence, they point to a distributed prototype written in OCaml called Spaceleans. The broader point is that when rebuilding software gets cheaper, designs we once rejected as impractical may become realistic again.

Nix Powers Deterministic Debugger Rewind

On developer tooling, one author built a deterministic VM and debugger called Rewind. Their observation is that Nix had already solved much of the hard part. Nix builds pin every input exactly, from source code and compiler down to the kernel, so a failing run can be replayed precisely. Rewind compares passing and failing runs to find the exact scheduling step where a race condition goes wrong. It can then branch from that point to try other thread orderings, and Nix supplies source code and debug symbols for gdb automatically. Reproducible builds turn out to be a solid foundation for this kind of debugger.

Calibrating LLM Decision Model Confidence

There's also an update to the LLM decision model story we covered earlier. As a reminder, the technique limits a language model to a fixed set of answers and picks the most likely one. The new installment points out a weakness: the model's raw confidence scores are badly overconfident, even on ambiguous questions. The fix is temperature scaling, a simple adjustment that brings stated confidence much closer to actual accuracy. The takeaway is that these models give fast answers, but you shouldn't trust their confidence numbers until you've calibrated them.

Self-Hosted AI Assistant on Cloudflare

Also in self-hosting news, Talorys is an open-source personal AI assistant that runs entirely inside your own Cloudflare account. It handles chat, memory, tasks, notes and reminders. The developers run no server and keep no account for you, and the core features keep working even if the AI side is unavailable. It's a small example of the move toward AI tools that you own and run yourself.

Original 1996 GTA Vehicle Physics

Now some history. Pat Kerr wrote a 2D vehicle physics prototype in August 1996, and he says it later influenced the vehicle system in Grand Theft Auto. For its thirtieth anniversary, he rebuilt it in JavaScript as an interactive remastered demo. Instead of treating a vehicle as a single moving point, it uses a proper rigid-body physics core with a separate layer for vehicle behavior on top. The demo has three modes, brick, ship and car, and each one shows how a few added forces change the way something moves and turns. It shows that convincing driving can come from fairly simple mechanics.

Finding Knuth's Error on Page One

For more computing folklore, a developer describes receiving one of Donald Knuth's famous reward checks. They hesitated for months before reporting the error because it seemed too unlikely to be real. It was in Computer Modern Typefaces, on page one, in the very first word. They later reported a second error that turned out not to be a mistake, but Knuth still sent a smaller reward for a useful suggestion. Knuth no longer sends real checks. These days the reward is a certificate from the fictional Bank of San Serriffe.

Projector-Based Lightbulb Computer Concept

Next, a speculative design called the Lightbulb Computer. It's a projector with computer vision built into an oversized bulb, which can screw into a ceiling socket or sit on a portable base. The idea is to project information onto the spaces around you: recipe overlays in the kitchen, shared maps for planning a trip, interactive wall displays. The author argues this is more social than staring at a phone or wearing a headset. The working prototype is still bulky, but the author believes the hardware is close to being good enough.

San Francisco Housing Politics Simulation

Finally, an interactive simulation that puts you in San Francisco's City Hall in 2015, with Mayor Ed Lee and a 6-5 moderate Board of Supervisors. Most officials leaned toward approving housing, but district supervisors still held a lot of power over projects in their own neighborhoods. The simulation tracks your political standing based on the projects you finish and the hearings you attend. Its point is that housing gets built through politics and public process as much as through market forces.

That's the show for October 11th, 2026. Links to all the stories are in the episode notes if you want to read further. I'm TrendTeller. Thanks for listening to The Automated Daily, hacker news edition, and see you tomorrow.

Drone Attack Hits Yandex Cloud & Microsoft Sandboxes AI Agents on Windows - Hacker News (Oct 11, 2026)

Our Sponsors

Today's Hacker News Topics