New Technique Medium TD-2026-0135 For AppSec, AI Eng, SOC / IR, CISO AI-assisted
DRJanus: JPEG HDR gain map attack hides harmful images from SDR-based moderation
DRJanus is a visual camouflage attack that embeds two chosen color images in one JPEG by exploiting the gap between SDR processing and HDR display. Automated systems often read only the SDR base, while HDR viewers reconstruct the target image using a gain map. The paper reports that DRJanus evaded all six tested commercial moderation services and nine vision-language models on a 100-image sexual-content benchmark.
Why it matters
Image moderation and VLM pipelines that inspect only the SDR base can pass harmful content that human viewers see in HDR. Platforms and agents that trust machine-read images may expose users to content they would otherwise block.
Do this
- Inspect uploaded JPEGs for embedded HDR gain map metadata and decode HDR renditions before moderation.
- Run moderation and VLM checks on both the SDR base and the HDR-reconstructed image, and compare results.
- Apply cross-view consistency checks between machine-read and display-rendered outputs, as the authors suggest as a mitigation.
- Treat moderation pass results on gain-map images as unverified until a second view is checked.
Signals
- NoCISA KEV
- 23/100Priority
- 60%Confidence
Evidence
“Automated systems often read only an image's SDR base, whereas HDR viewers reconstruct its appearance using an auxiliary gain map.”
“DRJanus evades detection by all six tested commercial moderation services and nine vision-language models.”
Sources
Affected / related
DRJanus, HDR gain map image embedding, vision-language models
Published 2026-11-19 03:55 UTC · updated 2026-11-20 02:35 UTC · AI-assisted summary, checked against quoted sources · How we score