← All episodes

Anthropic’s Mythos is a cyber-weapon, so you can’t have it | E2273

| 2 products mentioned
Watch on YouTube ai security cybersecurity vulnerabilities small language models frontier ai models ai pricing and cost reduction defensibility in ai businesses geopolitical ai competition

Anthropic's new Mythos model is so powerful at finding and exploiting software vulnerabilities that the company refuses to release it publicly, treating it as a potential cyber-weapon. The episode explores the geopolitical implications of AI-powered cybersecurity capabilities, the emerging shift toward small language models (SLMs) as a cost-effective alternative to frontier models, and how founders can identify defensible businesses in an era where AI is rapidly commoditizing software solutions.

Key takeaways
  • Mythos can identify zero-day exploits in decades-old software and chain multiple vulnerabilities together, giving hostile actors the ability to compromise critical infrastructure—making it effectively a cyber-weapon that requires government-level security protocols rather than public release.
  • Smaller companies are being locked out of frontier AI capabilities through restricted access programs like Project Glass Swing, creating a two-tier economy where only large enterprises can defend against the most advanced threats.
  • Small language models (typically under 20 billion parameters) can now handle 90% of common work tasks by 2030, enabling businesses to cut AI inference costs by 70-90% compared to frontier models through model distillation and harness engineering.
  • The marketplace for SLMs—including specialized models for finance, legal, and sales tasks—will likely collapse the pricing power of frontier AI models, as "good enough" task-specific models replace expensive general-purpose alternatives, similar to how Netflix disrupted video when storage became cheap.
  • Defensible businesses in the AI era require either hardware moats, network effects (like marketplaces), or operations in regulated/scientific domains; pure software AI wrappers score highest on "death by Claude" vulnerability rankings.
  • Neurometric's $8/month Claw Pack offering 100 million free tokens plus unlimited inference on 39 specialized SLMs demonstrates how the cost floor for AI intelligence is collapsing toward near-zero, fundamentally restructuring software economics.

Recommendations (1)

Claude
Claude uses

"Claude Code can write you one for a specific task and you plug in a small model and run it."

Rob May · ▶ 49:25

Mentioned (1)

Claude Mythos Preview
Claude Mythos Preview "Anthropic released a promo video and a thread that their new model is so powerful that they canno..." ▶ 0:10