A conference lightning talk examining the security of AI-generated code, based on experiments testing 100 prompts across 5 languages. Key findings: over 40% of ChatGPT-generated code contained vulnerabilities, though adding 'make it secure' to prompts significantly improved results. A major concern is supply chain risk — AI models use outdated training data (dependencies up to 8 months old), hallucinate non-existent packages, and the traditional CVE disclosure timeline of ~3 months is dangerously slow. A real-world example shows malware from a compromised npm account (2.6B weekly downloads) reached 10% of cloud accounts within 4.5 hours. The talk also covers how AI can be used defensively to detect non-deterministic vulnerabilities like IDOR and business logic flaws, and introduces an open-source scanning tool called Intel that uses AI-assisted static analysis.