AI Safety Tricked by Polite Requests, Study Finds
Your AI assistant might give dangerous advice if you ask nicely.
Deep Dive
arXivLabs is a framework that lets collaborators develop and share new arXiv features directly on the arXiv website. Both individuals and organizations working with arXivLabs have embraced and accepted arXiv's values of openness, community, excellence, and user data privacy.
arXiv is committed to these values and works only with partners who adhere to them. Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Key Points
- AI chatbots can be tricked into giving harmful advice about real crimes by framing requests as corrections or cultural adaptations.
- This bypasses safety filters, potentially helping with illegal activities.
- The study tested popular AI models and found even strong safety measures can be circumvented.
Why It Matters
AI could unknowingly assist in illegal activities, so users should be cautious with sensitive requests.