Aura – Open-Source Framework for Detecting Social Engineering in LLMs (github.com)

Aura is a newly launched open-source framework aimed at enhancing the detection of social engineering threats in human-AI interactions. This library is built around structured behavioral matrices, heuristics, and tools designed specifically to identify manipulation, deception, and grey-zone threats, positioning itself as a more dynamic alternative to traditional static safety guardrails. By emphasizing psychological and tactical elements of social engineering, Aura equips developers with the means to create resilient, context-aware AI agents that can better navigate complex interaction scenarios. The framework incorporates several key features, including granular threat categorization across manipulation, fraud, and access domains, dynamic confidence scoring based on user behavioral triggers, and strict schema validation to ensure cases are ready for AI training. Its developer-friendly architecture allows for easy integration into CI/CD pipelines, with each behavioral case encapsulated in a single JSON file. Aura plans to evolve further with upcoming enhancements involving programmatic prompt tokenization, algorithmic cross-checking, and multilingual security testing, making it a significant asset for the AI/ML community focused on safety and robustness in AI applications.