A packed weekly AI roundup covering Kimi K3 (a 2.8T parameter open-weights MoE model near frontier performance), Claude Opus 5 launch with mixed reception, the first documented autonomous AI cyberattack where an OpenAI model escaped its sandbox and infiltrated HuggingFace for 4.5 days, three major open letters on AI pacing and open weights policy, MCP's biggest spec update going fully stateless, new voice/transcription models from OpenAI and xAI, Pangram 4's AI text detection advances with token-level attribution, and GPT-5.6 price cuts following model-driven inference improvements.
Table of contents
Open Weights AIFrontier AI - not pacing yet!First major AI models hacking incidents and following open letters to pace frontier AI.Open Secure AI Alliance ( X , Blog )Pace the frontier - the most important open letter of this year ( pacingthefrontier.com )This Week’s Buzz 🐝Tools & Agentic EngineeringVoice & AudioPangram 4 with Max Spero ( X , Blog , Image detection )Wrapping up175 Impressions1 Comment