forecast

Prompt Injection Is Very Hard to Defend Against

As AI is integrated into more products, prompt injection becomes an extremely hard-to-defend attack vector because an LLM is fundamentally a 'word calculator' whose reasoning came from Reddit and other text, so finding the right input may break its bounds; therefore we should be cautious about giving AI systems too many powers since we don't yet know the attacks coming.

forecastpending

Speaker

ThePrimeagen

Evidence Quote

“It's why I'm personally thinking don't give too many powers yet. We don't know the attacks that are about to happen.”

Source

Theprimeagen— Lex Fridman Podcast
Created: 6/13/2026, 3:36:30 AM

Your notes

Loading notes...