#chatbots

How LLM Chatbots Leak Data Through Their Own Rendered Output

A recurring AI-security finding: an injected instruction makes the model emit a markdown image whose URL carries the user's data to an attacker server.