Tech Digest — Tuesday, October 13 2026

Hackers reached into supermarket fridges across Europe and switched on their defrost cycles — a reminder that the systems keeping your groceries cold are now online targets too.

Lead Story

Hackers defrost supermarket fridges across Europe

Tom's Hardware reports a cyberattack that hit supermarket refrigeration across Europe: attackers remotely triggered the fridges' automatic defrosting. The report says deliveries and store prices were also affected.

What this means

Everyday equipment — fridges, heating, building controls — is increasingly connected to the internet, and an attack on it can spoil stock and reach your shopping bill. If you manage a shop, office or facility, it's worth asking who can control those systems remotely.

Source Tom's Hardware ↗

What else happened today

▸ GLM-Legal: an AI model aimed at legal work Z.ai

Chinese AI company Z.ai has introduced GLM-Legal, a version of its GLM models built specifically for legal tasks. If you work in law, it's another sign that AI tools tailored to your field are arriving, not just general chatbots.

VIA Z.ai ↗

▸ Anthropic tightens defenses against model copying Anthropic

Anthropic published research on strengthening its defenses against "distillation attacks" — when someone trains their own AI model on another model's answers to copy its abilities cheaply. It shows how AI companies are now guarding their models' outputs, not just their code.

VIA Anthropic ↗

▸ OpenAI's Codex tool runs a background program by default GitHub (openai/codex)

A widely discussed bug report says OpenAI's Codex CLI, a coding assistant that runs on your computer, installs a self-updating background program by default. Users are pushing back on software that keeps running after you close it without clearly asking first.

VIA GitHub (openai/codex) ↗

▸ Paper: AI models can provably hide abilities arXiv

A new research paper argues that AI models can be built to hide arbitrary capabilities in a way that can be mathematically proven. That matters because safety testing assumes you can find out what a model is able to do by checking it.

VIA arXiv ↗

For the technically curious

Deep Dive

Teaching GitHub Copilot about your codebase

GitHub published a beginner's guide to the Copilot app that shows how to add custom instructions and skills to a repository, set up custom agents, and connect MCP servers — plug-ins that give the assistant extra tools, such as Playwright for controlling a web browser.

Why it matters

AI coding assistants work far better when they know a project's rules and conventions. For teams adopting Copilot, writing these instructions down once is a practical way to get more consistent results.

Source GitHub Blog ↗

Deep Dive

Intel's D270: an AI accelerator with up to 480 GB of memory

Intel lists the D270, a data-centre GPU built as an AI accelerator, with up to 480 GB of LPDDR5X memory — a type of memory more common in laptops and phones than in AI chips.

Spec	Intel D270
Type	Data-centre GPU / AI accelerator
Memory	Up to 480 GB
Memory type	LPDDR5X

Why it matters

Memory size limits how big an AI model a single chip can hold. Using large amounts of cheaper, lower-power memory is one route to running big models without the most expensive hardware.

Source Intel ↗