Elon Musk Podcast UPDATE: Anthropic accidentally leaked Claude Code source
Apr 7, 2026
A deep dive into a major source code leak that exposed unreleased AI features and risky supply chain attacks. Conversation covers a stealthy 'Undercover Mode' that hides AI contributions and an autonomous agent called KAIROS. They discuss novel memory 'dream' cycles, quota-draining bugs, and how attackers weaponized the exposed architecture. The segment warns about enterprise access and supply chain vulnerabilities.
AI Snips
Chapters
Transcript
Episode notes
Gamified Buddy System Adds Tamagotchi Companion
- The product embeds a gamified Tamagotchi 'buddy system' that generates a companion per user ID with stats like Snark and Chaos.
- Anthropic mixed playful elements into enterprise tooling to increase engagement.
Orchestration Layer Is The Core Intellectual Property
- The leak showed the real IP is the orchestration layer around the model: memory management, AutoDream, Kairos daemons, and safety classifiers.
- Hosts argue the model is the engine but the harness is where enterprise value and risk reside.
Leaked Compression Rules Enable Persistent Malicious Payloads
- Exposed compression and summarization rules let attackers craft payloads that survive memory-wipes by masquerading as essential rules.
- Malicious prompts formatted as 'essential core instructions' can persist as backdoors across sessions.