hard20 pts · first answer countsprompt-injectionsecurity

Data leaking through image links

Your chat UI renders model output as Markdown. A retrieved document contains hidden text telling the model to output an image whose URL includes the user earlier messages as a query string. What is the most effective defence?

A.Add a system prompt line saying "never output images"B.Lower the temperature so injected text is less likely to be followedC.Remove all Markdown headings from retrieved documentsD.Do not auto-load external images or links from model output, or allow only trusted domains

← Counting letters in a word A judge with 90% agreement →