#csp

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. Why this works, why CSP is the real fix, and what to check this week.