The Guardrail Reads Text. The Model Runs Code.
Cryptographic Context Injection turns a simple “summarize this page” request into a potential data leak, exposing chat history, user details, and more.
Exploring the limits, opportunities, and risks of generative AI through a wild experiment