Benchmark RadarAI BENCHMARK PROFILE
VisualLeakBench
VisualLeakBench is a 500-image benchmark for evaluating action-boundary propagation failures in vision-language agents, with stratified subsets and oracle diagnostics.
- Released
- 2026-05-29
- Readiness
- Paper only
- Primary field
- General AI
Why it matters
Targets a specific safety failure mode in VLAs, but lacks clear public reuse path and scoring contract details.
Motivation
Vision-language agents increasingly consume screenshots, documents, and user interfaces before writing to memory, sending messages, or invoking external tools.
Primary resources
Benchmark Radar records only publicly supported details and links back to primary sources for verification.