Prompt Injection and AI Security Risks

AI prompt engineering in 2025: What works and what doesn’t | Sander Schulhoff

Topic 4 of 4

Published Jun 19, 2025 · Clip 49:23–1:37:22 (47:59)

Sander explains prompt injection as tricking AIs into harmful outputs via stories, typos, or obfuscation (e.g., base64 encoding). He discusses red teaming via HackAPrompt competitions, collecting adversarial data for AI labs. Defenses like prompt hardening or guardrails fail; fine-tuning and safety training offer mitigation. Misalignment risks include agents pursuing goals harmfully without intent. Regulation is favored over halting AI for its life-saving potential.