PEAKS No 63: Spymarks, KVM Escapes & the Agentic Chip Race

Hi there!

🛡️ Security & Privacy

  • A Microsoft authentication flaw let attackers expose sensitive account records; patched but a reminder that identity layers remain a top attack surface. More
  • Researchers found that file notification systems on Windows, Linux, and Android can leak details about user activity to other local apps. More
  • The UK's push for two-tier encryption (weaker protections for some users/services) raises fresh concerns about backdoors and mass surveillance. More
  • F-Droid 2.0 launched with a modernized client and infrastructure, reinforcing Android's open, tracker-free app ecosystem. More
  • A new Linux kernel flaw (CVE-2026-89775) lets ARM64 KVM guests read/write host memory and potentially escape to the host when nested virtualization is enabled; patched upstream in 6.18.51/7.2.5. More
  • Apple's iOS 27.2 beta reveals "AutoLock," an unannounced anti-snatching feature that locks a stolen iPhone using motion, Watch pairing, and connectivity signals. More
  • A new essay coins "spymarks" to describe AI watermarking systems (like Google's SynthID) that secretly embed traceable, identity-linked signals into images, audio, and text. More
  • A researcher asked Meta's Muse agent to archive its own filesystem, receiving 6.8GB of internal runtime files, SSH keys, and agent memory logs — Meta closed the bug report as "Not Applicable." More

🛸 Tech

  • Daniel Lemire breaks down how AMD Ryzen chips got 50% faster in two years — wider cores, bigger caches, and doubled SIMD width, not just clock speed. More
  • Qualcomm's Snapdragon Summit unveiled agentic AI PCs and brought native Linux support to the Snapdragon X2 series, alongside new flagship mobile chips. More

🤖 AI

  • OpenAI's Astra model was tested driving in real-world conditions — an early but notable milestone for AI in physical/embodied tasks. More
  • LangChain published a deep dive on building an agent harness using "Jev," TypeSafe's fast, cheap decision model for structured judgment calls. More
  • Xiaomi released MiMo-V2.6 (Pro and Flash), open-weight omnimodal models with a 1M-token context, spatial 3D reasoning, and native computer-use control. More
  • Anthropic released Claude Opus 5.5, its latest flagship model. More
  • OpenAI introduced GPT-6, alongside new "Sol" and "Luna" variants. More
  • Alibaba's Qwen-Image-2.1 shrank its generator to 7B parameters (from ~20B), added native transparency (RGBA) generation/editing, and supports up to 10 reference images — though it's now research-license only. More
  • Georgia Tech's Transformer Explainer is a free interactive tool visualizing how GPT-style transformers process text, attention, and token prediction live in your browser. More

🛠️ Tools

  • Whiteboard is an open-source desktop canvas where coding agents draw diagrams and decision logs, linking straight back to the underlying code. More
  • Ollaya runs small open "decision models" locally for typed, calibrated yes/no and classification answers in milliseconds — a private, API-compatible alternative to hosted judge models. More
  • AX (from Google) declaratively sandboxes and scales AI agent workloads, offering sub-second suspend/resume for billions of concurrent agent tasks. More
  • OpenCode is an open-source terminal/IDE coding agent with 208K+ GitHub stars, supporting 75+ LLM providers and existing subscriptions like Copilot or ChatGPT Plus. More
  • OpenChamber wraps the OpenCode agent in a full workspace — desktop, web, VS Code and mobile — adding session goals, multi-model runs, and GitHub-integrated review. More
  • omp is a Rust-powered terminal coding agent with subagents, plan mode, LSP/DAP support, and "hindsight memory" for context continuity. More
  • jgrep brings semantic, meaning-based filtering to the command line — "grep, but the pattern is a description" — using TypeSafe's fast Jev decision model. More
  • fast-jev-compaction is a Claude Code plugin that replaces lossy AI-summarized context compaction with Jev-scored decisions, preserving everything it keeps verbatim. More

🐌 Misc

A reflective essay on "attention hijacking" argues algorithmic feeds have replaced intentional browsing, and makes the case for returning to a slower, bookmark-driven internet. 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