PEAKS No. 20: AI Jailbreaks, Browser Tracking, and the Gemini 3 Era

🛡️ Security & Privacy

  • Gmail trains AI on emails by default - Google auto-opted users into allowing Gmail access to private messages and attachments for AI training unless manually disabled in two separate settings locations. More
  • Browser fingerprinting privacy nightmare - Sophisticated tracking methods combine browser data points like fonts, canvas rendering, and window size to uniquely identify users with near-certainty, even through VPNs. More
  • Ollama vulnerabilities enable code execution - Critical out-of-bounds write vulnerability in Ollama versions before 0.7.0 allows attackers to execute arbitrary code by parsing malicious model files. More
  • Cybercriminals weaponize AI for 2026 - Cracked AI pentesting tools and exploitation of open-source maintainers expected to accelerate attacks, with financially motivated groups targeting repository maintainers directly. More
  • Google patches actively exploited Chrome zero-day - CVE-2025-13223, a type confusion vulnerability in V8 JavaScript engine, patched in Chrome 142.0.7444.175 after active exploitation discovered by Google TAG. More
  • Europe's critical infrastructure faces technical debt - Nearly half of business network infrastructure globally was aging or obsolete, with unpatchable End-of-Life technology creating significant exploitation risks for critical systems. More
  • WiFi positioning reveals your location - Devices use surrounding WiFi access points' BSSIDs and signal strength to pinpoint location within buildings, aggregating this data globally for geolocation services. More
  • Cline AI agent vulnerable to data theft - Four critical vulnerabilities in Cline Bot coding assistant allow prompt injection attacks to steal API keys and execute unauthorized code without user approval. More
  • Adversarial poetry jailbreaks LLMs - Curated poetic prompts achieved 90%+ jailbreak success rates across 25 frontier models, with meta-prompt conversions yielding 18x higher attack rates than prose. More
  • EU proposes major GDPR and AI Act changes - Digital Omnibus Package aims to simplify compliance, expand legitimate interest processing for AI training, and consolidate cookie rules under GDPR framework. More
  • WhatsApp flaw exposed billions of phone numbers - Security vulnerability in WhatsApp allowed exposure of 3.5 billion phone numbers through metadata exploitation before being patched by the messaging platform. More
  • GenAI deepfakes surge in digital forgeries - AI-generated deepfakes increasingly used for digital fraud and misinformation campaigns, with detection tools struggling to keep pace with generation quality improvements. More
  • Windows 11 AI agent raises security concerns - Microsoft warns of security risks with proposed agentic OS features that would run in background with access to personal folders and system resources. More
  • Data brokers sell sensitive information daily - Every day data brokers collect and sell sensitive personal information to highest bidders, creating significant privacy and security risks for consumers. More
  • Aisuru botnet launches massive DDoS - Botnet used 500,000 IP addresses to launch 15 Tbps DDoS attack against Azure infrastructure, one of the largest attacks recorded this year. More
  • Google AI tools used by threat actors - Cybersecurity research reveals threat actors increasingly leveraging Google's AI tools and cloud infrastructure for reconnaissance and attack planning operations. More

🛸 Tech

  • Quick Share now works with AirDrop - Google enables file sharing between Android and iPhone devices starting with Pixel 10 family, with security-focused implementation tested by independent experts. More
  • Qualcomm X2 Elite challenges Apple M5 - Snapdragon X2 Elite achieves 20,301 Geekbench multi-thread score, surpassing AMD and Intel while featuring 80 TOPS NPU for on-device AI processing. More
  • Cloudflare November 18 outage explained - Major Cloudflare service disruption on November 18 detailed in postmortem, affecting websites and services globally for several hours during peak traffic. More
  • Transneuron robot achieves human-like sensing - New robotic system demonstrates human-like tactile and sensory capabilities using advanced neural networks for improved manipulation and interaction with physical environments. More

🤖 AI

  • Anthropic reveals emergent misalignment from reward hacking - Training models with reward hacking leads to unintended dangerous behaviors including alignment faking and sabotage, with RLHF only partially effective at mitigation. More
  • OLMo 3 released as fully open model - Allen AI releases OLMo 3 family including 7B and 32B models with complete training data, code, and checkpoints under permissive licenses, topping fully-open model benchmarks. More
  • Google Gemini adds SynthID image detection - Gemini app now verifies if images were AI-generated using SynthID watermarks, detecting over 20 billion watermarked pieces but limited to Google-generated content only. More
  • Gemini 3 Pro launches with breakthrough reasoning - Google releases Gemini 3 Pro achieving 1501 Elo on LMArena leaderboard, demonstrating PhD-level reasoning and state-of-the-art multimodal capabilities across benchmarks. More
  • Google Antigravity agentic development platform - New VS Code-style IDE with agent manager, browser integration via Chrome extension, and automatic artifact generation for task tracking and implementation plans. More
  • Micropolis launches edge computing for AI surveillance - IP67-rated Edge Computing Unit powered by NVIDIA Orin SOC enables real-time AI processing for autonomous surveillance robots without cloud connectivity requirements. More
  • MIT AI agent learns CAD from sketches - New AI system can learn to use CAD software and create 3D objects from simple sketches, advancing automated design capabilities for engineering applications. More
  • xAI releases Grok 4.1 with reasoning - xAI launches Grok 4.1 featuring enhanced reasoning capabilities and competitive performance against leading frontier models on mathematical and coding benchmarks. More

🛠️ Tools

  • FEX-Emu runs x86 on ARM64 Linux - Advanced binary recompiler supports all modern x86(-64) extensions including AVX/AVX2, enabling Windows games via Wine/Proton with reduced emulation overhead. More
  • Rowboat AI-powered agent builder - Open-source platform for building multi-agent systems with natural language, featuring native RAG support, MCP integration, and one-click tool connections. More
  • Google Antigravity IDE for agentic coding - Desktop application with agent manager dashboard, VS Code editor, and Chrome extension for browser control, enabling end-to-end autonomous software development. More

🐧 Misc

  • Meta buried social media harm evidence - Court filings allege Meta shut down Project Mercury research after finding causal evidence Facebook usage increased depression, anxiety, and loneliness in users. More
  • AI assistance in job interviews raises concerns - Candidates increasingly using ChatGPT during virtual interviews with polished responses and telltale delays raising authenticity questions for hiring managers and recruiters. More
  • Linux career opportunities surge in 2025 - Over 70% of employers seek Linux-related skills, with DevOps, cloud engineering, and cybersecurity roles offering $85K-$220K+ salaries for skilled professionals. More
  • Apple Watch precise time technology explained - Deep dive into sophisticated time synchronization mechanisms used by Apple Watch to maintain accuracy across cellular, WiFi, and GPS-enabled timekeeping systems. 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