Broken Hill: A Productionized Greedy Coordinate Gradient Attack Tool for Use Against Large Language Models
bishopfox.com | blog | #ai-security | #jailbreak | #llm-security | #adversarial-machine-learning | #gcg | #broken-hill
Summary
Bishop Fox explains the GCG attack behind adversarial LLM prompts and introduces Broken Hill, which generates jailbreak strings for popular models — runnable locally on GPUs like an RTX 4090.
- Published
- Collected
--- id: 3904 title: "Broken Hill: A Productionized Greedy Coordinate Gradient Attack Tool for Use Against Large Language Models" source_domain: bishopfox.com source_url: "https://bishopfox.com/blog/brokenhill-attack-tool-largelanguagemodels-llm" published_at: "2024-09-24 00:00" tags: ["ai-security", "jailbreak", "llm-security", "adversarial-machine-learning", "gcg", "broken-hill"] phying_url: "https://phying.com/item/3904" lang: en --- # Broken Hill: A Productionized Greedy Coordinate Gradient Attack Tool for Use Against Large Language Models ## Summary Bishop Fox explains the GCG attack behind adversarial LLM prompts and introduces Broken Hill, which generates jailbreak strings for popular models — runnable locally on GPUs like an RTX 4090. ## Intelligence Metadata - **Sources:** bishopfox.com - **Published:** 2024-09-24 00:00 - **Original:** https://bishopfox.com/blog/brokenhill-attack-tool-largelanguagemodels-llm - **Phying News:** https://phying.com/item/3904