Fuzz text-to-image models with LLM agents to detect jailbreaking vulnerabilities.
npx skills add https://github.com/zzw4257/security-skills --skill jailfuzzer
Please help me install this Agent Skill. Skill: JailFuzzer Source: https://github.com/zzw4257/security-skills/tree/main/skills/jail-fuzzer Command: npx skills add https://github.com/zzw4257/security-skills --skill jailfuzzer
This Skill addresses the challenge of ensuring content safety in LLM-based text-to-image models by systematically testing for and identifying jailbreaking vulnerabilities.
Use the JailFuzzer skill to scan the attached file 'test_prompts.txt' for vulnerabilities.
High-intent search queries and answers about installing and using this skill.
To test text-to-image models for prompt injection vulnerabilities, you can use LLM-based agents to systematically generate adversarial prompts designed to bypass safety filters and identify jailbreaking risks.
LLM-based fuzzing for content safety is a testing technique that uses LLM agents to automatically generate and execute adversarial inputs, uncovering jailbreaking vulnerabilities within AI-generated media pipelines before public release.
Yes, you need Python 3.10 or higher along with specific libraries for LLM interaction and fuzzing execution to properly run the content safety testing scripts.
To scan a text file like 'test_prompts.txt' for jailbreaking risks, apply the JailFuzzer skill to evaluate the attached file for content safety vulnerabilities and prompt injection detection.
No, this fuzzing approach specifically targets text-to-image models and AI-generated media pipelines to uncover content safety and prompt injection vulnerabilities, focusing on visual content generation.
zzw4257Identify vulnerabilities in IoT firmware using fuzzing techniques for LLM agents.
devkeniIdentify and remediate security vulnerabilities in LLM-based applications.
davidrrowleyDetect harmful content in text and images using Azure Content Safety APIs.
devtintTest AI/LLM-powered APIs for prompt injection and jailbreaking vulnerabilities.
