Marketplace All-in-One Bytes: Week in Review — Anthropic's new AI model, a referendum on data centers, and NASA livestreams journey to space
Apr 10, 2026
Joanna Stern, technology columnist and founder of New Things, discusses Anthropic’s Claude Mythos Preview and its limited Project Glasswing access. She covers why AI companies hold back powerful models and the tradeoffs that creates. Joanna also explores a Wisconsin referendum pushing back on data centers and how astronauts use social media and NASA livestreams to bring space closer to the public.
AI Snips
Chapters
Transcript
Episode notes
Powerful Security-Finding Model Held Back
- Anthropic's Claude Mythos Preview finds deep security vulnerabilities so effectively the company is limiting public release.
- Anthropic formed Project Glasswing to give access to 40+ companies like Google and JPMorgan Chase for responsible testing and patching.
Self-Editing Becomes Default For Sensitive Models
- AI companies are increasingly self-editing models and restricting releases due to misuse risks like surveillance or weapons creation.
- Joanna Stern compares this to companies gating powerful models to researchers and partner firms before public deployment.
Cybersecurity Becomes Human Plus AI
- Cybersecurity work will shift to humans managing AI agents that surface vulnerabilities rather than replacing experts.
- Joanna Stern likens the change to coding and customer service where humans instruct AI and verify results.