Daily Arxiv

This is a page that curates AI-related papers published worldwide.
All content here is summarized using Google Gemini and operated on a non-profit basis.
Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.

A Survey on the Safety and Security Threats of Computer-Using Agents: JARVIS or Ultron?

Created by
  • Haebom

Author

Ada Chen, Yongjiang Wu, Junyuan Zhang, Jingyu Xiao, Shu Yang, Jen-tse Huang, Kun Wang, Wenxuan Wang, Shuai Wang

Outline

This paper presents a knowledge systematization of safety and security threats to computer-assisted agents (CUAs). While CUAs, evolving into LLM-based systems, can autonomously manipulate desktop applications, web pages, and mobile apps, they pose new safety and security risks due to the vulnerabilities of LLM inference and the complexity of integrating diverse software components and multimodal inputs. Through a literature review, this paper presents a definition suitable for safety analysis of CUAs, a classification of current safety threats, a comprehensive classification of defense strategies, and benchmarks, datasets, and evaluation metrics used to evaluate the safety and performance of CUAs. This provides guidance for future research and the design and deployment of secure CUAs.

Takeaways, Limitations

•
Takeaways:
◦
Provides a systematic understanding of CUA's safety and security threats.
◦
Systematically classifies CUA's security threats and suggests defense strategies.
◦
Summarizes benchmarks, datasets, and evaluation metrics for CUA safety and performance evaluation.
◦
It provides a structural foundation for future research and actionable guidance for practitioners.
•
Limitations:
◦
Further research is needed to determine how well the classifications and benchmarks presented in this paper can adapt to new threats and technological advancements in the future.
◦
It focuses on general threats and defense strategies rather than specific vulnerability analysis of specific CUA systems or applications.
◦
There is a lack of experimental verification of the effectiveness of the proposed defense strategy.
👍
Made with Slashpage