ai_security▌

← feed

NOTABLE

arXiv: MobileProbe — screen-content injections hijack mobile GUI agents, chosen step by trajectory

Mobile GUI agents act on whatever the screen shows, so instructions embedded in ordinary UI content can redirect their actions (environmental injection). MobileProbe drops fixed payload-at-a-step benchmarks and instead picks injection steps trajectory-aware, then adaptively evaluates agent security across models.

released 2026-11-12 · ingested 2026-11-13T12:20Z

Sources

Primary source

Cite this

APA

AI Security Index. (2026). arXiv: MobileProbe — screen-content injections hijack mobile GUI agents, chosen step by trajectory. AI Security Index. https://arxiv.org/abs/2611.16693

BibTeX

@misc{aisec-2026-11-13-arxiv-2611-16693-mobileprobe-environmental-injection-mobile-agents,
  title  = {arXiv: MobileProbe — screen-content injections hijack mobile GUI agents, chosen step by trajectory},
  year   = {2026},
  howpublished = {\url{https://arxiv.org/abs/2611.16693}},
  note   = {AI Security Index item 2026-11-13-arxiv-2611-16693-mobileprobe-environmental-injection-mobile-agents}
}