How AI is changing code security

Code security for software engineers

Topic 11 of 14

Published Nov 26, 2025 · Clip 51:36–58:22 (6:46)

Johannes explains AI's impact on security tooling: AI can enrich deterministic analyzers, e.g., gathering knowledge about millions of libraries to feed taint analysis, and is very effective at fixing specific, deterministically found issues in small contexts rather than whole codebases. He describes how architecture changes with LLMs in the backend introduce prompt injection—text becomes code and prompt injection is the new code injection. With coding assistants, writing code is no longer the challenge; verifying ever-faster-produced code at scale and speed is the new bottleneck, and he points to training LLMs on datasets free of security issues as a pioneering direction. Finally, he notes AI makes the same mistakes as humans but shifts prevalence—slop squatting (AI inventing non-existent packages that attackers register) accelerates, while one-liner issues like hardcoded passwords may decline, and complex multi-snippet issues may rise.