Artemis II crew splashes down safely, France moves to replace Windows with Linux, and Sam Altman responds to Molotov cocktail attack on his home

TechCrunch

France to Ditch Windows for Linux to Reduce Reliance on US Tech

France is planning a government-wide transition away from Microsoft Windows toward open-source Linux operating systems as part of a broader push for digital sovereignty and reduced dependence on American technology companies. The move reflects growing European concern about strategic exposure to foreign corporate control over critical government infrastructure, accelerated by the current US administration's geopolitical posture and trade uncertainty. France joins a wave of European nations — including Germany's ongoing migration to open-source software — treating software stack independence as a national security issue rather than a procurement preference.

Sam Altman's Blog

Sam Altman's Response to Molotov Cocktail Attack on His Home

Sam Altman posted a personal reflection following an attack in which a Molotov cocktail was thrown at his home, partly attributing the climate of hostility to inflammatory media coverage around AI development. He reaffirmed his conviction that AI represents humanity's greatest opportunity for expanding human capability, while acknowledging legitimate safety concerns and admitting past mistakes — including conflict-aversion and board management failures at OpenAI. Altman called for de-escalation of adversarial rhetoric, arguing that the right path forward is democratising access to AI and establishing democratic oversight, rather than concentrating corporate control or resorting to confrontation. A suspect has since been arrested.

Linux Kernel Documentation

Linux Kernel Formalises Policy on AI Coding Assistants

The Linux kernel project has added formal guidelines for contributors using AI coding assistants, choosing transparency over prohibition. The policy prohibits AI agents from adding "Signed-off-by" certifications — only human developers may legally certify code under the Developer Certificate of Origin, preserving human accountability for every patch. Contributors must instead disclose AI assistance using a new "Assisted-by" tag specifying the tool name and model version. The guidelines reflect a pragmatic position: AI tools are welcome in the development workflow, but humans remain the responsible parties who must understand, review, and vouch for every line submitted to the kernel.

Eclectic Light

Why You Can't Trust macOS Privacy & Security Settings

A detailed investigation has uncovered that macOS's Privacy & Security panel displays inaccurate information about application access to protected folders such as Documents and Desktop. A test application called Insent demonstrates that once an app gains access by a user selecting a folder through an Open/Save dialog, macOS sandboxing constraints are bypassed and full access persists — even after the permission is toggled off in Settings. The flaw stems from a disconnect between the sandboxing layer and the TCC privacy database operating at different levels; revoking access requires terminal commands and a system restart, steps no ordinary user would know to take. The finding means macOS users cannot rely on the Privacy & Security panel to accurately reflect what their applications can actually read and write.

Cloudflare Blog

Cloudflare Reaches 500 Tbps of Global Network Capacity

Cloudflare announced that its global network has crossed 500 terabits per second of external interconnection capacity — the sum of every port facing transit providers, peering partners, and Internet exchanges across its 330+ cities — a milestone that reflects 16 years of compounding infrastructure investment. In 2025, the network autonomously mitigated a 31.4 Tbps DDoS attack in 35 seconds without paging a single engineer, made possible by a fully distributed defence architecture in which every server independently evaluates and drops attack traffic at line rate via eBPF before packets can reach application layers. The post also notes that AI crawlers now account for more than 4% of all HTML requests across Cloudflare's network — comparable to Googlebot — and grew over 15× in 2025, prompting new TLS fingerprinting and behavioural analysis techniques to distinguish legitimate crawlers from attacks.

WireGuard Mailing List

WireGuard Releases First Windows Update Since Microsoft Signing Dispute

WireGuard for Windows has released its first update since a Microsoft code-signing account suspension briefly halted the project's ability to ship signed drivers — an issue that drew significant attention alongside a simultaneous suspension affecting VeraCrypt. The new release, WireGuardNT v0.11 and WireGuard for Windows v0.6, includes accumulated bug fixes, performance improvements, and the ability to remove individual allowed IPs without disrupting active connections. Project lead Jason Donenfeld clarified that the suspension was a routine bureaucratic delay resolved within a day, dispelling speculation about a broader Microsoft policy against security tooling. The episode underscores the structural fragility of open-source security software that depends on big-tech platform accounts to remain usable on modern Windows.

Simon Willison's Weblog

ChatGPT Voice Mode Runs on a Much Weaker Model Than Users Expect

Simon Willison flags a non-obvious capability gap in OpenAI's product line: voice mode operates on a GPT-4o era model with an April 2024 knowledge cutoff, while the company's top-tier coding tools have made dramatic strides through reinforcement learning on verifiable tasks. The counter-intuitive implication is that the interface which feels most capable — conversational voice — is backed by the weakest model, while the products that look most narrowly technical have received the deepest investment because their results can be objectively verified. Willison draws on Andrej Karpathy's observation that this creates a growing perception gap: users who encounter AI primarily through voice assistants and social media will have a fundamentally different and far less accurate model of AI capability than users working with code or research tools.

Stripe Engineering Blog

Selective Test Execution at Stripe: Fast CI for a 50M-Line Ruby Monorepo

Stripe describes how it built a dynamic test selection system for a 50-million-line Ruby monorepo with 1.2 million test cases — a codebase where static dependency analysis fails because of Ruby's pervasive use of metaprogramming. Rather than trying to infer which tests cover which code, Stripe built a C++ interceptor library that hooks into the filesystem at the OS level during test execution, recording exactly which files each test actually reads at runtime and storing the result as a compressed bitmap index. On an average CI run, the system now executes just 5% of the full suite — a median of 0.5% of tests — while consuming less than 10% of previous compute, with guardrails that fall back to full runs when dynamic patterns such as directory globbing are detected.

Construction Physics

Helium Is Hard to Replace — and the Supply Chain Is Under Pressure

A detailed look at helium's supply chain reveals that the element is both uniquely irreplaceable and dangerously concentrated: no substitute exists for its 4.2 Kelvin boiling point, which is essential for MRI superconducting magnets, EUV semiconductor lithography, fibre optic manufacturing, and aerospace applications. Supply has historically been dominated by the US and Qatar; the US Strategic Helium Reserve was fully depleted in 2024, and current Middle East tensions have disrupted Qatari exports. With semiconductor demand projected to require five times the current helium supply by 2035, the article argues that helium has quietly become one of the most strategically significant materials in the global technology stack — without attracting the policy attention given to rare earth metals or lithium.