PEAKS No 52: GLM 5.2 vs Claude in Cyber Benchmarks, EU Chat Control Fight Returns, Lab-Grown Cell Divides
Hi there!
Last week I started reading “A Philosophy of Software Design” by John Ousterhout. It feels like the right book at the right moment. The main subject is how to deal with complexity in code: how to identify, track and fix it. Some principles sound like common sense, but that common sense only develops after you've spent real time working with high-quality code. Having everything presented in a structured manner really helps to grasp all the concepts and apply them intentionally. These principles matter even more when you're reviewing AI-generated code.
🛡️ Security & Privacy
- Patrick Breyer warns of undemocratic backroom deals reviving the EU's Chat Control proposal, prompting relaunch of the fightchatcontrol.eu campaign. More
- Security researchers disclosed three AirDrop vulnerabilities affecting Apple devices; Apple is developing a comprehensive fix. More
- A massive, ongoing password spray campaign is targeting Azure CLI accounts, prompting warnings from security researchers. More
- A year-old unpatched vulnerability reportedly renders Apple's Hide My Email feature ineffective at protecting user identities. More
- F-Droid published a nuanced take on how "malware" gets defined and flagged in the Android app ecosystem. More
🛸 Tech
An engineer documents building an FMCW radar startup to detect asbestos in walls, using Capon beamforming and a CNN to classify material signatures. More
🤖 AI
- Semgrep's cyber benchmarks show GLM 5.2 outperforming Claude models on real-world security tasks, raising questions about frontier AI safety leaderboards. More
- OpenClaw brings AI agents to Android and iOS, letting users run autonomous assistants directly from their phones. More
- Researchers found Claude Code silently embeds steganographic markers into requests, raising transparency and trust questions for developers. More
- Anthropic launched Claude Science beta, a specialized tool aimed at accelerating scientific research workflows. More
- Google DeepMind released Nano Banana 2 Lite (Gemini Flash-Lite image model), a lightweight image-generation option for faster, cheaper output. More
- Mistral released Leanstral 1.5, a free Apache-2.0 Lean 4 proof-engineering model that saturates miniF2F and solves 587 of 672 PutnamBench problems, even uncovering real code bugs. More
- Quesma argues Qwen 3.6 27B hits the sweet spot for local AI development, balancing performance and hardware requirements. More
- Proton's Lumo 2.0 adds image generation, persistent memory, and private web search to its privacy-first AI assistant. More
- X (Twitter) launched an MCP server, making its platform natively accessible to AI agents and tools. More
- Anthropic shipped a Claude Desktop beta for Linux, bringing Chat, Cowork, and Claude Code with parallel sessions, visual diff review, and an integrated terminal to Debian-based distros. More
🛠️ Tools
- Google's Copybara automates transforming and syncing code changes between different repositories and version control systems. More
- PeerTube offers a decentralized, ActivityPub-federated video streaming platform with peer-to-peer playback in the browser. More
- SearXNG is a free, self-hostable metasearch engine that aggregates results without tracking or profiling users. More
🧬 Misc
- A veteran engineer outlines five habits that separate world-class software engineers from everyone else. More
- Scientists built the first fully synthetic cell capable of growing and dividing, a milestone for synthetic biology. More
📩 Please feel free to share this article with colleagues and friends who will find it valuable.
Thanks for reading!
Have a great day!
Bogdan