CompressionAttack: Exploiting Prompt Compression as a New Attack Surface in LLM-Powered Agents
Published 17 Nov 2025 · arXiv · Zesen Liu
Overview
The paper presents CompressionAttack, a framework that exploits prompt compression in large language models (LLMs) as a new attack surface. This vulnerability arises because compression modules prioritize efficiency over security, allowing adversarial inputs to alter LLM behavior.
Key Insights
- CompressionAttack Framework: Utilizes HardCom and SoftCom strategies to manipulate LLMs.
- Attack Success Rate: Achieves up to 83% and 87% success rates in different tasks.
- Real-World Impact: Case studies demonstrate practical implications, with current defenses proving ineffective.
BFSI Relevance
- Why Relevant: LLMs are increasingly used in BFSI for automation and decision-making.
- Primary Sector: Financial Services
- Subsectors: AI-driven decision-making, fraud detection
- Actionable Implications: BFSI professionals should enhance security protocols for AI systems and consider alternative compression strategies.
researcher peer-reviewed-paper global