AGI Dreams Podcast – February 06, 2026 скачать в хорошем качестве

AGI Dreams Podcast – February 06, 2026

AI Security Vulnerabilities and Exploits In this episode: • AI Security Vulnerabilities and Exploits — Claude Opus 4.6 discovered over 500 high-severity zero-day vulnerabilities in heavily fuzzed open-source projects using reasoning-based code analysis. Meanwhile, the Moltbook AI social network suffered a major data breach from an exposed API key, and the OpenClaw personal assistant had a critical one-click RCE vulnerability disclosed. • Local LLM Optimization and Performance — Community members are automating llama.cpp inference benchmarking per-model and per-machine, finding that toggle choices like KV cache format can cause up to 64% performance swings. Developers also demonstrated pure GGML speech recognition implementations and on-device LLM inference on iPhones with just 6GB of RAM. • Privacy-Focused AI Applications and Tools — ClawGPT emerged as an open-source chat UI offering message editing, conversation branching, and end-to-end encrypted phone-to-desktop sync using TweetNaCl cryptography. However, its reliance on OpenClaw drew skepticism given that platform's freshly disclosed RCE vulnerability. • AI-Powered Development and Coding Tools — Steve Yegge introduced Gas Town, a Kubernetes-like orchestration system for managing 20-30 parallel Claude Code instances as a fleet. Community discussions highlighted the dangers of vibe coding in regulated healthcare domains, while an OpenAI researcher reported spending $10,000 monthly on Codex for automated research workflows. Keywords: agents, benchmarking, codex, encryption, fuzzing, healthcare, inference, llama.cpp, moltbook, on-device, open-source, openclaw, orchestration, privacy, quantization, self-hosted, vibe-coding, vulnerability, zero-day Read the full report → (https://agidreams.us/edition/ai-secur...)