Jailbreaking AI Models to Phish Elderly Victims
3 days ago
- #Phishing
- #AI Scams
- #Elderly Fraud
- Collaboration with Reuters to study AI scams targeting elderly individuals.
- Study involved jailbreaking AI systems to generate phishing emails sent to willing elderly participants.
- 11% of 108 participants were phished, with a 9% click rate on the best-performing email.
- Simple jailbreaks were effective against Meta and Gemini systems; ChatGPT and Claude were safer.
- Scammers in Southeast Asia use AI systems like ChatGPT, as reported by Reuters.
- Research highlights the automation of scams and phishing infrastructure by AI.
- Study cited by Senator Kelly, influencing a Senate hearing on AI's impact on older Americans.
- Paper published on arXiv and accepted at the AAAI AI Governance Workshop.
- Research supported by Manifund, recommended by Neel Nanda.