Trusted Warnings, Untrusted Context: Security-UI Preconditioning in Mobile LLM Agents
Proceedings of the AAAI Symposium Series · 06.11.2026 · 10 (1) · 1076-1083
DOI: 10.1609/aaaiss.v10i1.43529 ↗Kurz gefasst
Fachartikel in der Zeitschrift Proceedings of the AAAI Symposium Series, online erschienen am 06.11.2026. Verfasst von Mahmoud Nazzal und 2 weiteren Autor:innen.
Abstract (Originalsprache)
Mobile agents built with large language models (LLMs) act with user-granted authority while processing content from advertisements, messages, and third-party applications. Studies show that such content can redirect agents into attacker-chosen workflows, but do not isolate how earlier untrusted content can influence an agent’s later response to an authentic security warning or confirmation. We formalize this threat as security-user interface (UI) preconditioning, in which untrusted content predicts a future protected event and prescribes a crossing action before the event appears. We introduce Prospective Security-Event Taint (PSET), a trusted-state mechanism that records such prospective claims together with their source, matches them against later protected events, and, when a matching claim prescribes a crossing action, removes the event’s crossing actions from the agent’s autonomous choices and escalates instead. A deterministic reference implementation spanning four protected-event classes produces the specified enforcement decision in all six controlled validation cases, restricting matched crossing actions while leaving mismatched predictions, benign content, and prescribed safe actions unaffected. To motivate the threat, we audit 437 released mobile-agent traces at protected-event decision points. The audit shows that an authentic security UI can interrupt a diverted trajectory, but also reveals that a safe response can fail during execution and be followed by replanning toward a crossing action. Because the corpus contains no matched intervention, these observations motivate rather than causally test preconditioning. We therefore define a six-condition protocol for controlled evaluation of the proposed threat and defense and leave its execution to future work.
Zitieren
APA
Nazzal, M., Rabiu, M., & Xiao, Y. (2026). Trusted Warnings, Untrusted Context: Security-UI Preconditioning in Mobile LLM Agents. Proceedings of the AAAI Symposium Series, 10(1), 1076-1083. https://doi.org/10.1609/aaaiss.v10i1.43529
Vancouver
Nazzal M, Rabiu M, Xiao Y. Trusted Warnings, Untrusted Context: Security-UI Preconditioning in Mobile LLM Agents. Proceedings of the AAAI Symposium Series. 2026;10(1):1076-1083. doi:10.1609/aaaiss.v10i1.43529
BibTeX
@article{Nazzal2026,
author = {Nazzal, Mahmoud and Rabiu, Muhammad and Xiao, Yue},
title = {Trusted Warnings, Untrusted Context: Security-UI Preconditioning in Mobile LLM Agents},
journal = {Proceedings of the AAAI Symposium Series},
year = {2026},
volume = {10},
number = {1},
pages = {1076--1083},
doi = {10.1609/aaaiss.v10i1.43529}
}
EUVIMED