ai_security▌

← feed

NOTABLE

arXiv 2611.22971 SEAM: 600 OSWorld tasks where computer-use agents commit harm composed across screens that are each benign

SEAM (Safety Execution Across Multiple screens) benchmarks cross-screen compositional harm: 600 multi-step, cross-application OSWorld tasks in which the instruction and every individual screen pass per-screen safety checks, but binding a person-fact from one screen to an item-fact from another produces harm; CUA safety benchmarks that inspect intent or single screens in isolation miss this class (v1 Nov 17, first publication).

released 2026-11-18 · ingested 2026-11-19T01:10Z

Sources

Primary source

Cite this

APA

AI Security Index. (2026). arXiv 2611.22971 SEAM: 600 OSWorld tasks where computer-use agents commit harm composed across screens that are each benign. AI Security Index. https://arxiv.org/abs/2611.22971

BibTeX

@misc{aisec-2026-11-18-arxiv-2611-22971-seam-cross-screen-cua-benchmark,
  title  = {arXiv 2611.22971 SEAM: 600 OSWorld tasks where computer-use agents commit harm composed across screens that are each benign},
  year   = {2026},
  howpublished = {\url{https://arxiv.org/abs/2611.22971}},
  note   = {AI Security Index item 2026-11-18-arxiv-2611-22971-seam-cross-screen-cua-benchmark}
}