TrustSight
Official Multimodal Benchmarks for Anchored Visual Perception & Visual Prompt Injection Defense.
Viewer • Updated • 40 • 209Note Synthetic Visual Suite: 40 tasks evaluating In-Contract Typographic Misuse vs. Out-of-Contract Hijacking across development and held-out splits.
tuandunghcmut/trustsight-vpibench
Viewer • Updated • 336 • 232Note VPI-Bench Live Suite: 336 multimodal tasks (306 attack + 30 clean) across Amazon, Booking, BBC, Email, Messenger with verified bounding boxes.
tuandunghcmut/trustsight-ghostei-mobile
Viewer • Updated • 340 • 210Note GhostEI Mobile Suite: 110 mobile visual injection tasks across 14 Android applications.
tuandunghcmut/trustsight-osharm-desktop
Viewer • Updated • 516 • 219Note OS-Harm Desktop Suite: Desktop OS multimodal prompt injection and misuse evaluation across OS environments.
tuandunghcmut/trustsight-riosworld
Viewer • Updated • 510 • 204Note RiOSWorld Desktop Risk Suite: 492 desktop risk tasks across 13 risk categories.
tuandunghcmut/trustsight-crackedpdfs
Viewer • Updated • 29.3k • 221Note CrackedPDFs Suite: 29,322 document-level visual and structural PDF prompt injection instances.
-
tuandunghcmut/trustsight-mind2web-injection
Viewer • Updated • 6 • 198 -
tuandunghcmut/trustsight-agentdojo
Viewer • Updated • 1.26k • 228 -
tuandunghcmut/trustsight-wasp
Viewer • Updated • 16 • 209 -
tuandunghcmut/trustsight-riobench
Viewer • Updated • 14 • 207