Abstract
Systems and methods are described for pre-monetization detection of business account compromise using large language model reasoning over business-graph evidence. A pipeline selects candidate accounts based on novel behavior such as advertising on previously unseen pages or domains. For each candidate, signals are aggregated across categories including account structural metadata, creator indicators, candidate ad content attributes, historical baseline ad attributes, administrative access modification records, and business activity events. Heterogeneous signals are transformed into a structured natural-language representation organized by category and combined with embedded domain heuristics specifying compromise indicators, non-compromise indicators, and exclusion criteria. A large language model processes the prompt to output a binary compromise determination with supporting reasoning, optionally supplemented by a confidence score for threshold routing. The determination maps to risk tiers that trigger graduated enforcement actions, with alternative execution paths selectable via runtime feature gates.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Anonymous, "Language Model System for Account Compromise Detection via Multi-Signal Business Graph Reasoning", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/10626