Anthropic's new model Mythos found a 27-year-old bug in OpenBSD and a 16-year-old bug in FFmpeg given nothing more than the source code and an open-ended prompt. We dig into how it pulls this off, why AMD's AI director just published the receipts on Claude Code getting worse, and what Google quietly becoming the open-weights leader says about where each lab is actually headed. Plus: Spark's weird new RL compression trick, and why every model release this week is really a strategy memo in disguise.