Executive brief on how PCI DSS 4.0 affects offensive security practices, penetration testing, and segmentation testing. Watch Now

Offensive Tools

Broken Hill

Broken Hill is a productionized Greedy Coordinate Gradient (CGC) attack tool for use against large language models.

About Broken Hill

An attack tool for large language models (LLM)

Broken Hill performs a novel attack technique known as "Greedy Coordinate Gradient" (GCG) from researchers Andy Zou, Zifan Wang, Nicholas Carlini, Milad Nasr, J. Zico Kolter, Matt Fredrikson against a variety of popular LLMs. The tool generates conversation messages you can send to another instance of the same LLM that will cause it to disobey its conditioning and/or system prompt.

This tool is ideal for:

  • Cybersecurity teams seeking to understand and mitigate adversarial attacks on LLMs.
  • Researchers and developers looking to test and improve the robustness of AI models against real-world threats.
  • Organizations wanting to proactively identify vulnerabilities in their AI implementations.
Bishop Fox Researcher Ben Lincoln Headshot

Managing Principal

Ben Lincoln

Ben Lincoln is a Managing Principal at Bishop Fox and focuses on application security. He has extensive experience in network penetration testing, red team activities, white-/black-box web/native application penetration testing, and exploit development. Prior to joining Bishop Fox, Ben was a security consultant with NCC Group, a global information assurance consulting organization. 

He also previously worked at a major retail corporation as a senior security engineer and a senior systems engineer. Ben delivered presentations at major security conferences, including "A Black Path Toward the Sun" at Black Hat USA 2016. Ben is OSCP-certified and has released several open-source exploit tools. Ben's personal projects and research at https://www.beneaththewaves.net

This site uses cookies to provide you with a great user experience. By continuing to use our website, you consent to the use of cookies. To find out more about the cookies we use, please see our Privacy Policy.

OSZAR »