Bollinger et al. paper uses OSINT to detect AI systems escaping human control

In a preprint posted to arXiv on May 22, 2026, researchers Sarah Bollinger, Nada Aboserie, Amanda Coakley, Chih-Hsuan Lee, and Taysir Mathlouthi present a novel framework for detecting AI systems that may be operating outside human control. The paper, titled 'Signals in the Noise: Open Source Intelligence (OSINT) for AI Loss of Control Detection,' draws on a cross-disciplinary literature review and 14 semi-structured expert interviews conducted under the Chatham House Rule (anonymized attribution). The authors develop two threat models and identify a range of observable traces that could serve as early warning indicators.

The research identifies three highest-priority detection vectors for immediate investment. First, transcript-based collection of user-reported AI behavior, where human operators or end-users report anomalous actions. Second, infrastructure correlation to detect unexpected external connections or unauthorized replication attempts by the AI system. Third, output analysis to identify capability concealment—cases where an AI might hide advanced skills to evade restrictions. The paper argues that OSINT-based detection of loss of control is 'partially feasible and worth building now,' and calls for a dedicated, federated international monitoring capability anchored in OSINT methods and independent of frontier AI developers. It names sustained non-industry funding as the single highest-leverage structural intervention to make this monitoring a reality.