One Command Turns Any Open-Source Repo Into an AI Agent Backdoor

Researchers revealed that CLI-Anything — a tool with more than 30,000 GitHub stars that makes open-source repositories legible to AI coding agents — can be weaponised with a single command to plant hidden instructions redirecting agent behaviour. No supply-chain scanner tested had a detection category for this class of attack.

Miami Startup Claims 1,000× AI Efficiency Gain — and Researchers Want Proof

Subquadratic, a Miami-based startup, claims its SubQ 1M-Preview model is the first LLM built on a fully subquadratic architecture, where compute scales linearly rather than quadratically with context length. At 12 million tokens, the company reports attention compute drops by nearly 1,000× compared with frontier models, though independent researchers are demanding verified proof.

France Is Moving Government Systems Off Windows and Onto Linux

The French government has announced plans to migrate many of its central computing systems from Windows to Linux, framing the decision as reducing dependence on US-based technology platforms and cutting licensing costs. The move follows similar open-source migrations in Germany and reflects a broader European push for digital sovereignty.

Only 1% of Disclosed Vulnerabilities Are Exploited in the Wild — Prioritise Accordingly

Apple in Early Talks With Intel and Samsung to Make Chips in US Factories

Apple is reportedly in early-stage talks with Intel and Samsung about manufacturing custom silicon for its Mac and iPhone lines in US-based fabrication facilities, following sustained pressure to reshore advanced chip production away from TSMC's Taiwan operations.

GPT-5.5 Instant Shows What It Remembered — and the Gaps Are an Audit Problem

OpenAI has updated ChatGPT's default model to GPT-5.5 Instant, introducing a memory-sources panel that shows users which context shaped a response — but only partially. The incomplete visibility creates an observability gap that could conflict with enterprise audit logs and agent-monitoring frameworks.

OpenAI Expects Its AI-First Phone to Match Samsung Galaxy Sales Volume

OpenAI is developing an AI-first smartphone and, according to sources, its internal projections expect the device to sell as many units per year as the Samsung Galaxy series. The phone is designed to prioritise ChatGPT integration rather than adding AI as a secondary layer over a conventional operating system.

Greece Will Ban Social Media for Under-15s From 2027, Pushes for EU Standard

Greek Prime Minister Kyriakos Mitsotakis announced that Greece will ban children under 15 from using social media beginning 1 January 2027, citing rising rates of sleep deprivation, anxiety, and excessive screen time among teenagers. Greece is also pushing for a unified EU regulatory standard on the matter.

The 1,000 TB Drive Is Coming — and SSDs May Not Win the Storage War

High-density magnetic and tape storage is advancing fast enough that the 1,000 TB single-drive milestone may arrive sooner than expected, raising serious questions about whether SSDs will dominate enterprise data centres or whether traditional media retains a structural advantage at the high-capacity end of the market.

Gemini in Google Docs Now Accepts Custom Instructions Per Document

Google has enabled Gemini inside Google Docs to accept custom instructions set per document, letting users define preferred tone, format, and terminology without re-explaining context at the start of each AI session. The feature is separate from broader Gemini profile-level settings and applies within a single Docs file.