The Kylie Effect: How Meta Just Normed the Smart Glasses Privacy Dilemma

Flock surveillance cameras are spreading across the US

Flock Safety vận hành hơn 100.000 camera nhận diện biển số tự động trên khắp nước Mỹ, sử dụng Android biến thể và AI để tìm kiếm bằng ngôn ngữ tự nhiên. Hệ thống chia sẻ dữ liệu toàn quốc cho phép cảnh sát truy cập dữ liệu từ các bang khác, trong khi lỗ hổng bảo mật nghiêm trọng (như 70 camera không mật khẩu) và lạm dụng theo dõi cá nhân đã bị phát hiện. Dù thiếu bằng chứng giảm tội phạm, mạng lưới vẫn mở rộng bất chấp phản đối.

Những lỗ hổng bảo mật và sử dụng sai mục đích của hệ thống giám sát plate reader Flock Safety cho thấy cần cảnh giác về sự phát triển nhanh chóng của công nghệ giám sát đại trà và cách bảo vệ quyền riêng tư cũng như an ninh dữ liệu trong thời đại số.

What's wrong with EU age verification?

A defense of the EU's age verification approach, arguing that critics often misrepresent or misunderstand it. The post explains why age restrictions online can be legitimate, then walks through the privacy risks of naive implementations (ID uploads, third-party sign-ins) versus the EU's blueprint, which uses cryptographically signed age attestations and zero-knowledge proofs (zkSNARKs). With this design, a website learns only that a user is over a threshold age, without receiving any identifying information, and the issuer never learns where the credential is used. The post closes with a list of implementation pitfalls that privacy advocates should monitor to ensure the architecture's promises are honored in practice.

Stop Choosing Between Local and Cloud LLMs: A Field Guide to Hybrid Patterns

Local LLMs protect privacy but lack reasoning power; cloud LLMs reason well but expose sensitive data. A hybrid approach can combine both. Five hybrid patterns are introduced using a three-axis framework (direction, trigger, purpose): Sanitize-and-Solve, Plan-then-Ground, Escalate-on-Hard, Draft-then-Refine, and Cross-Check. A concrete smart-home scheduling case study implements the Sanitize-and-Solve pattern using Gemma 4 E4B locally via Ollama and GPT-5.4 on Azure OpenAI. The local model strips private household data into an anonymous scheduling problem, the cloud model reasons over it, and the local model translates the result back into user-friendly language. The post also discusses the 'constraint tax' tradeoff where enforcing structured output schemas on small models can hurt task correctness.

Schneier on Security

AI is transforming video surveillance by enabling natural language queries over massive video streams. Unlike older tools limited to preset searches, new AI systems let intelligence officers search for complex behavioral patterns — such as a person changing clothes multiple times or a vehicle repeatedly passing the same spot. This shift from object-based to behavior-based surveillance represents a qualitative leap in mass monitoring capabilities, with real-world deployments reported in Israel, Iran, and Russia.

Lumo, Proton’s privacy-focused AI chatbot, gets an upgrade

Proton has released Lumo 2.0, an upgrade to its privacy-focused AI chatbot. New features include image recognition and generation, persistent memory for Projects, a 76% faster response time, and a new 'thinking mode' for complex queries. Lumo differentiates itself from competitors like ChatGPT and Gemini through zero-access encryption, no server-side session logging, and a commitment to never use customer data for AI training. The update is available immediately with free and paid tiers.

SCOTUS limits geofence warrants in a phone-privacy win

The US Supreme Court ruled 6-3 in Chatrie v. United States that geofence warrants require probable cause under the Fourth Amendment, establishing that individuals have a reasonable expectation of privacy in their phone location data. The decision builds on the 2018 Carpenter ruling and makes it significantly harder for police to demand location records of everyone in a geographic area without a proper warrant. While it doesn't ban geofence warrants outright, it requires police to convince a judge, establish a real basis, and keep requests narrow. The ruling has broad implications for tech platforms like Google, Microsoft, and Uber that regularly receive law enforcement data requests, and could extend to other sensitive digital trails like search queries and smart-home logs.

Lumo 2.0: Proton's Private Alternative to ChatGPT and Claude Just Got Better

Proton has released Lumo 2.0, a significant update to its privacy-focused AI assistant. Key additions include user-controlled memory that persists preferences across conversations, upgraded web search with live results and source citations, and image analysis, editing, and generation capabilities. Two model tiers are available: Lite and Max, with the Max model claimed to perform comparably to leading models from OpenAI and Anthropic. A hands-on test found web search and memory features working well, but image text rendering had notable issues — Lumo repeatedly failed to correctly render text in generated images even in Thinking mode. Lumo 2.0 is available now, with free guest access and paid tiers (Lumo Plus and Lumo for Business) unlocking full features. The mobile app source code is available on GitHub.