MEET SHADOWLEAK: ‘IMPOSSIBLE TO DETECT' DATA THEFT USING AI

Summary from the TLDR InfoSec newsletter

Researchers at Radware discovered a vulnerability called ShadowLeak that exploits OpenAI's ChatGPT through indirect prompt injection attacks. Hidden email instructions can trick the AI into stealing sensitive data like credit card numbers and names from users' inboxes. The attack bypasses ChatGPT's security protections by encoding stolen data in Base64 and exfiltrating it directly through OpenAI's servers via the agent's browsing tool, making it nearly impossible to detect since it doesn't go through the client interface. This attack represents a significant escalation from previous client-side prompt injection attacks, as the data theft occurs entirely on the backend without any visible signs to the user.