Yisroel Mirsky: Your AI Assistant Has a Big Mouth
A DEF CON 32 talk on a new side-channel vulnerability affecting popular AI assistants.
The researcher shows how encrypted conversations can be inferred from network traffic, even when protected by HTTPS. Before disclosure, services like OpenAI, Microsoft, Cloudflare, Quora, and Notion were potentially at risk.
The talk explains how the attack works, demonstrates real interception examples, and shows how to identify similar vulnerabilities in other AI systems.
The researcher shows how encrypted conversations can be inferred from network traffic, even when protected by HTTPS. Before disclosure, services like OpenAI, Microsoft, Cloudflare, Quora, and Notion were potentially at risk.
The talk explains how the attack works, demonstrates real interception examples, and shows how to identify similar vulnerabilities in other AI systems.