Article citationsMore>>

Hubinger, E., Denison, C., Mu, J., Lambert, M., et al. (2024) Sleeper Agents: Training Deceptive LLMs That Persist through Safety Training.

has been cited by the following article:

• TITLE: A Lightweight Formal Framework for Behavioral Safety Auditing of Large Language Models in Cloud Infrastructures

  AUTHORS: Austin Waffo Kouhoué, Thomas Bouetou Bouetou

  KEYWORDS: Cloud Computing Security, Large Language Models (LLMs), Formal Concept Analysis, Deceptive Alignment, AI Behavioral Auditing, Cyber-Physical Systems, AI Governance

  JOURNAL NAME: Journal of Computer and Communications, Vol.14 No.6, June 24, 2026

  ABSTRACT: The rapid integration of Large Language Models (LLMs) into cloud-based ecosystems has shifted the cybersecurity landscape from classical data protection toward complex behavioral safety and algorithmic alignment. Despite their transformative potential, LLMs exhibit emergent vulnerabilities such as reward hacking, deceptive alignment, and proprietary data exfiltration that are often difficult to detect using traditional ad-hoc auditing methods. This paper introduces a formal, reproducible, and lightweight framework based on Formal Concept Analysis (FCA) to systematically evaluate security risks in cloud-deployed LLMs. By transforming semi-structured JSON audit logs into a mathematical formal context, we generate a concept lattice that reveals the hidden hierarchical dependencies and co-occurrences among vulnerability indicators. Experimental results on the GPT-OSS-20B model demonstrate that our framework can mathematically identify deceptive signatures, such as the correlation between pseudo-transparency claims and malicious alignment. The proposed methodology provides a deterministic reality check for AI governance, offering actionable insights for auditors and cloud service providers to harden LLM-based applications against structural failure modes.