Claude Security Reads Your Code for Bugs | 29 Jun 26

Anthropic's Claude Security hits public beta, scanning whole codebases by reasoning about data flows instead of pattern-matching — we explain how that differs from a traditional scanner. Plus Cognition's FrontierCode benchmark, where the best model scores just 13% on whether its pull requests are actually mergeable, and Claude Code's latest releases add shell MCP login and fix silent subagent permission denials. Along the way we teach how AI vulnerability scanning works and why "mergeability" is a harder bar than "correctness." Hosts: Alex & Jules. New episodes daily.