Defensive LLMs Evaluated Against AI Social Engineering.

Yuqiao Xu, Osama Zafar, Alexander Nemecek, Erman Ayday· August 12, 2026 View original

Key takeaways

  • Defensive LLMs need to localize structural risks, not just react to surface cues.
  • Protective action and correct structural localization are often decoupled.
  • Asset-control failures are a major bottleneck for defensive LLM localization.
  • Live interaction testing is crucial for evaluating defensive AI effectiveness.

Who benefits

CybersecurityBFSIE-commerceSocial MediaGovernment

Summary

This research evaluates defensive LLMs against AI-generated social engineering attacks in live, turn-by-turn interactions, focusing on their ability to localize the structural source of risk rather than just surface cues. The study found that while models avoided unsafe compliance, protective action and correct structural localization were often decoupled, highlighting the need for more sophisticated defensive capabilities.

New research delves into the effectiveness of defensive large language models (LLMs) in combating AI-generated social engineering attacks during live, interactive scenarios. The study moved beyond simple detection, formalizing "trust-chain localization" to assess if defenders could identify the root structural failure of an attack, such as issues with actor authority or asset control, rather than just reacting to superficial cues. A controlled corpus of 300 online housing scenarios, encompassing legitimate cases and various structural failure modes, was used to evaluate five different defender models in both stateful turn-by-turn and one-shot static settings. While none of the models explicitly complied with unsafe requests, their defensive effectiveness varied significantly, with intervention rates ranging from 0% to 96.3%. A critical finding was the frequent decoupling of protective action from correct structural localization; models sometimes intervened for the wrong reason or recognized a structural failure without taking protective action. Asset-control failures proved particularly challenging for localization, and surface sensitivity varied across models, underscoring that safe-looking behavior alone is insufficient for robust scam resistance.

Why it matters

As AI-generated social engineering becomes more sophisticated, developing LLM-based defenders that can accurately identify and mitigate structural risks in real-time is crucial for protecting users and systems.

How to implement this in your domain

  1. 1Develop defensive LLMs that prioritize structural risk localization (e.g., actor authority, asset control) over mere surface-level detection.
  2. 2Design evaluation frameworks that separately measure intervention, timing, structural localization, and false-positive rates for defensive AI.
  3. 3Integrate specific training data and fine-tuning objectives to improve defensive LLMs' ability to identify asset-control failures.
  4. 4Conduct live, turn-by-turn interaction testing for defensive AI to capture real-world performance nuances.

Original post by Yuqiao Xu, Osama Zafar, Alexander Nemecek, Erman Ayday

"arXiv:2608.10239v1 Announce Type: new Abstract: Generative AI makes social-engineering attacks more fluent, adaptive, and scalable, increasing the need for LLM-based de- fenders that can protect users during ongoing interactions. We ask whether such defenders identify the structu…"

View on X

Originally posted by Yuqiao Xu, Osama Zafar, Alexander Nemecek, Erman Ayday on X · view source

Want to go deeper?

Turn these trends into skills with Learnijoy's hands-on AI & tech courses.

Explore courses

More in AI Research

AI Engineering & DevToolsAI Research

TACTICL Compresses Tabular ICL Models, Retaining Adaptability.

TACTICL is an automated framework for compressing tabular in-context learning (ICL) models by jointly pruning transformer layers and replacing them with lightweight adapters. This method significantly reduces model size and computational demands while preserving robustness to data shifts and in-context adaptability.

Mykhailo Koshil, Matthias Feurer, Katharina EggenspergerAug 12, 2026
AI Engineering & DevToolsAI Research

MoE Proxy Models Cut LLM RL Debugging Costs.

This paper introduces Mixture-of-Experts (MoE) proxy models designed for low-cost reproduction and diagnosis of failures during Large Language Model (LLM) Reinforcement Learning (RL) post-training. These proxy models significantly reduce computational resources and time needed for debugging, while accurately preserving training dynamics and fault responses.

Yikai Wang, Chuansai Zhou, Yuhang Zhou, Weiqiang Wu, Cong Wu, Yue Deng, Ben Feng, Mingming Zhu, Beirong Zhou, Zhibin Wang, Sheng Zhong, Chen Tian, Wangze ZhangAug 12, 2026
AI Engineering & DevToolsAI Research

New Algorithm Boosts Stochastic Optimal Control Efficiency.

This paper introduces Path Integral Value Matching (PI-VM), a novel value-based algorithm for Linear Quadratic Stochastic Optimal Control (LQ-SOC) that significantly improves computational efficiency and stability. By deriving a temporal recursive form of the value function and integrating Girsanov theorem with experience replay, PI-VM matches state-of-the-art precision with order-of-magnitude efficiency gains.

Bangyan Liao, Chenglei Yu, Yuchen Yang, Chuanrui Wang, Zhisheng Song, Peidong Liu, Tailin WuAug 12, 2026