AI News: The Scariest AI Model Ever!

TL;DR

Anthropic's unreleased Claude Mythos found vulnerabilities in major OS and browsers, so they're giving limited cybersecurity access via Project Glass Wing instead of public release.

Key Points

• 1.Claude Mythos is Anthropic's most powerful and dangerous model ever built. It scored 83.1% on cybersecurity vulnerability reproduction (vs. 66.6% for Opus 4.6) and beat Opus 4.6 by 24 points on SWEBench Pro — autonomously finding a 27-year-old OpenBSD flaw and chaining Linux kernel vulnerabilities.
• 2.Anthropic refused public release and launched Project Glass Wing instead. A select group of major tech companies' cybersecurity teams (not named OpenAI) get Mythos access to patch vulnerabilities before similarly powerful models reach bad actors.
• 3.The 'too dangerous to release' narrative has a marketing history. GPT-2 in 2019 generated apocalyptic headlines but ultimately proved manageable; the presenter believes Anthropic's fear is genuine this time, though acknowledges a partial marketing element.
• 4.Meta's Muse Spark is the first model from Meta Superintelligence Labs. It leaped from last place (Llama 4 Maverick) to fourth on the Artificial Analysis Intelligence Index, excelling at figure understanding but falling below Opus and Gemini on coding benchmarks — and it is not open source.
• 5.GLM 5.1 from ZAI is the overlooked story of the week. This MIT-licensed open-weight model scores 58.4 on SWEBench Pro, beating GPT-5.4 (57.7) and Opus 4.6 (57.3), meaning open-source models have effectively caught up to frontier closed models for coding.
• 6.Seed Dance 2.0 is now available in the US via Runway and CapCut. It effectively replaces Sora (recently discontinued), though viral features involving celebrity/IP generation have been removed; HeyGen Avatar 5 also launched this week, cloning identity from just 15 seconds of footage.
• 7.Anthropic cut third-party API usage from Claude subscriptions starting April 4th. Users can no longer use their $200/month Claude Max plan to power tools like OpenClaw, forcing users to switch to direct API keys or alternative providers like OpenAI.
• 8.A mystery model called Happy Horse 1.0 topped AI video leaderboards. Believed to be from Alibaba, it ranked above Seed Dance 2.0 and Kling; Google also quietly launched an offline iOS dictation app (Google AI Edge) using its Gemma model, and Spotify added AI-powered podcast discovery.