Bypassing LLM Defenses with Adversarial Suffixes and AutoDAN
summary
The talk explains what adversarial (competing) suffixes are and demonstrates that any large language model can be vulnerable to them. It shows how to construct such a suffix yourself and discusses why OpenAI does not regard these suffix-based attacks as a security threat.
AI-written summary · about
chapters (9)
- 0:00Introduction
- 1:08Adversarial attacks
- 3:58Generating adversarial suffixes
- 11:23Suffix transferability
- 13:54Defense methods
- 19:55AutoDAN features
- 25:43Attack examples
- 27:28Security testing frameworks
- 34:58Q&A
related talks
50mResponsible AI: Adversarial Attacks on LLMs
23mSecurity and auditing tools in Large Language Models (LLM)
01h 04mRobust Generalization in the Era of LLMs: Jailbreaking Attacks and Defenses
21mTricking Neural Networks : Explore Adversarial Attacks
31mHacking with Words...
01h 01mIntroduction to AI Security
49mSecuring AI Apps with the OWASP Top Ten for Large Language Models
48mStay Ahead of Adversarial AI in OT/ICS Environments – Mitigating CWE-1039