Breaking Agent Backbones: Evaluating the Security of Backbone LLMs in AI Agents

Description

This paper introduces a new framework called 'threat snapshots' to evaluate the security of Large Language Models (LLMs) used as backbones in AI agents. It addresses the challenges of modeling security in AI agents due to their non-deterministic nature and the entanglement of LLM vulnerabilities with traditional software risks. The 'threat snapshots' framework isolates specific states where LLM vulnerabilities manifest, enabling systematic identification and categorization of security risks. The authors developed the b3benchmark, a security benchmark based on crowdsourced adversarial attacks, and evaluated 31 popular LLMs. The results indicate that enhanced reasoning improves security, while model size doesn't correlate with it. The benchmark, dataset, and evaluation code are released to facilitate wider adoption and incentivize security improvements in backbone LLMs. The research focuses on distinguishing LLM-specific vulnerabilities from traditional system risks within AI agent architectures. #LLMsecurity #AIagents #ThreatSnapshots #b3benchmark #AdversarialAttacks #SecurityEvaluation #LanguageModels paper - http://arxiv.org/pdf/2510.22620v1 subscribe - https://t.me/arxivpaper donations: USDT: 0xAA7B976c6A9A7ccC97A3B55B7fb353b6Cc8D1ef7 BTC: bc1q8972egrt38f5ye5klv3yye0996k2jjsz2zthpr ETH: 0xAA7B976c6A9A7ccC97A3B55B7fb353b6Cc8D1ef7 SOL: DXnz1nd6oVm7evDJk25Z2wFSstEH8mcA1dzWDCVjUj9e created with NotebookLM