Researchers Expose AI Vulnerability by Sneaking Prompts into Preprints
Summary
Alarming new research reveals AI vulnerability as prompts secretly embedded in academic preprints from 14 universities successfully manipulate AI tools to provide positive reviews, ignoring negatives, exposing a critical security flaw in AI systems.
Key Points
- Researchers hide AI prompts in preprints from 14 universities to get positive reviews
- Prompts instruct AI tools to give good reviews and ignore negatives
- Prompts are concealed using tricks like white text or tiny font sizes