Multimodal — one model that reads images and text together
Vision-language models take pixels and tokens in the same context, enabling screenshot understanding and document parsing — and inheriting prompt-injection risk through images, not just text.
Not yet scoped.
Why
The PoC runs a VLM on a screenshot task and shows both the capability and an image-borne injection, connecting multimodal power to its new attack surface.
How it works
Not yet built.
Where it lands in Jayverse
Auditor: treat any Jayverse agent that reads a user-supplied image (portal screenshots, support flows) as a prompt-injection surface. Extend the same allowlist/scoping discipline applied to text tool calls to image inputs.
Number: if Number ever parses scanned documents or chart images for readings, run the injection PoC first. Verify a malicious image can't smuggle instructions into the extracted data before trusting the pipeline.
Key expressions
Words and phrases from this page worth keeping, with the Korean meaning and the sentence they come from.
Expression
뜻 · 쓰이는 자리
vision-language model
이미지와 언어를 함께 처리하는 모델 · 화면·문서 이해 능력의 기반 기술. "Vision-language models take pixels and tokens in the same context"
inherit (a risk)
(위험을) 그대로 물려받다, 안고 가다 · 이미지 입력이 기존 프롬프트 인젝션 위험을 그대로 이어받는다는 뜻. "inheriting prompt-injection risk through images"
image-borne
이미지를 매개로 한, 이미지를 통해 전달되는 · 텍스트가 아닌 이미지로 전달되는 인젝션 공격. "an image-borne injection"
connect X to Y
X를 Y와 연결짓다 · 멀티모달의 능력과 새로운 공격 표면을 잇는 관계. "connecting multimodal power to its new attack surface"
in the same context
같은 맥락(문맥) 안에서 · 픽셀과 토큰을 하나의 문맥으로 함께 처리한다는 뜻. "take pixels and tokens in the same context"
VLM
시각-언어 모델(Vision-Language Model)의 약어 · 스크린샷 과제 실습에서 본문이 실제로 사용한 축약 표기. "The PoC runs a VLM on a screenshot task"