← Voltar ao blog
Ameaças IAMédia

Image-based disclosure: Why the chat interface needs security review

An image address can carry information, but disclosure requires both embedding the information and making the external request.

Revisão editorial:

O texto deste artigo está disponível em inglês nesta versão do site.

Contexto e fontes

OWASP describes an indirect-injection scenario in which an assistant includes an image URL containing conversation data. If an interface requests that address, its contents can reach the destination server. Rendering and network policy therefore matter alongside generated text. Markdown alone does not prove that any request occurred. [1]

Âmbito e limites

Some interfaces block external content. An image request does not automatically read arbitrary files or other browser information.

Explicação simples

A hypothetical letter contains an image address with information written into it. That information leaves only if the address is actually contacted.

Recomendações práticas

We recommend testing external resources with fictional data while observing real network requests. Document permitted destinations and do not substitute a visual warning for technical restrictions. Include previews, exports and email views: the same model response may be handled differently by different interfaces.

Fonte primária / referência

  1. [1] OWASP LLM01:2025: Prompt Injection