[SSTF2024] Invited Talk 1 : AI Agents for Offensive Security скачать в хорошем качестве

[SSTF2024] Invited Talk 1 : AI Agents for Offensive Security

AI Agents for Offensive Security Brendan Dolan-Gavitt | Professor, New York University The rise of AI systems that can capably use tools and take actions in digital or physical environments promises to drastically change many fields, and computer security is no exception. In this keynote, I will discuss recent advances that have allowed language models to become surprisingly capable at autonomously identifying and exploiting software vulnerabilities. Building from the insight that the weaknesses of language models (hallucinations, non-determinism) can be avoided—or even turned into strengths—when strong external verifiers are available, I will show how offensive security is uniquely positioned to benefit from advances in AI. Finally, I will discuss novel pitfalls, open questions, and possible risks from the emergence of highly capable offensive security agents.