AI News: The Scariest AI Model Ever!

A weekly AI news roundup covering several major developments. The headline story is Anthropic's Claude Mythos, an unreleased frontier model with exceptional cybersecurity capabilities — finding decades-old vulnerabilities in OpenBSD, FFmpeg, and the Linux kernel — which Anthropic is restricting to select security firms via Project Glass Wing rather than public release. Other notable items include Meta's Muse Spark (first model from Meta Super Intelligence Labs, competitive but not best-in-class), GLM 5.1 (an open-weight MIT-licensed model matching GPT and Claude on coding benchmarks), Gemini's new interactive simulations and Notebooks features, Seed Dance 2.0 video model launching in the US via Runway and CapCut, HeyGen Avatar 5, OpenAI's new $100/month Codex tier, Anthropic ending Claude subscription coverage for third-party tools like OpenClaw, Cursor's remote agent control, and a mysterious top-ranked video model called Happy Horse 1.0 attributed to Alibaba.