# Use of LLMs for Illicit Purposes: Threats, Prevention Measures, and Vulnerabilities

**Authors:** Maximilian Mozes, Xuanli He, Bennett Kleinberg, Lewis D. Griffin  
**arXiv:** [2308.12833](https://arxiv.org/abs/2308.12833)  
**Date:** 2023-08-24  
**Summary page:** https://www.summarizepaper.com/en/paper/2308.12833v1/

## Key points

- The paper provides an overview of efforts to identify and mitigate security threats and vulnerabilities associated with large language models (LLMs).
- Weidinger et al. (2022) present a taxonomy of 21 risks associated with LLMs, including discrimination, misinformation harms, and malicious uses.
- Huang et al. (2023) categorize LLM vulnerabilities into inherent issues, intended attacks, and unintended bugs.
- Fan et al. (2023) focus on trustworthiness aspects of LLMs related to privacy, security, responsibility, and fairness.
- Bommasani et al. (2021) discuss the opportunities and risks of foundation models like BERT, CLIP, and GPT-3 in terms of technological aspects, societal impacts, legal consequences, and ethical issues.
- Kreps et al. (2022) examine the credibility of LLM-generated content compared to actual news articles.
- The summary emphasizes the need for awareness among developers and users regarding security-related problems associated with LLMs.
- It highlights existing works on LLM security concerns including potential criminal activities and the threat posed by credible LLM-generated misinformation if perceived as genuine by users.
- Limitations of prevention strategies are discussed along with potential future concerns arising from advancements in LLM development in terms of public perception.

## Summary

The paper provides an overview of existing efforts to identify and mitigate the security threats and vulnerabilities associated with large language models (LLMs). Weidinger et al. (2022) present a taxonomy of 21 risks associated with LLMs, categorizing them into areas such as discrimination, misinformation harms, malicious uses, and more. Huang et al. (2023) categorize LLM vulnerabilities into inherent issues, intended attacks, and unintended bugs. Fan et al. (2023) focus on trustworthiness aspects of LLMs related to privacy, security, responsibility, and fairness. Bommasani et al. (2021) discuss the opportunities and risks of foundation models like BERT, CLIP, and GPT-3 in terms of technological aspects, societal impacts, legal consequences, and ethical issues. Additionally, Kreps et al. (2022) examine the credibility of LLM-generated content compared to actual news articles. This refined summary emphasizes the need for awareness among developers and users regarding security-related problems associated with LLMs while providing an up-to-date presentation of existing works on LLM security concerns including potential criminal activities as well as the potential threat posed by credible LLM-generated misinformation if it is perceived as genuine by users. Limitations of prevention strategies are also discussed along with potential future concerns arising from advancements in LLM development in terms of public perception.