Keynotes
Title: Achieving a Secure AI Agent Ecosystem
Nicole Nichols, Distinguished Engineer @ Palo Alto Networks
Nicole Nichols is a Distinguished Engineer in Machine Learning Security at Palo Alto Networks. She previously held senior roles at Apple, Microsoft and has contributed to both academia and industry advancements in adversarial machine learning and security. She has published at numerous ACM, IEEE, and CVPR workshops, and was co-chair of ICML-ML4Cyber workshop. She has a PhD in Electrical Engineering from the University of Washington.
AI agent systems, capable of complex planning and autonomous action in real world environments, present profound and novel cybersecurity challenges. Current cybersecurity paradigms are too brittle to address the unique vulnerabilities stemming from dynamic generative agents with opaque interpretability, new protocols connecting tools and data, and the unpredictable dynamics of multi agent interactions. Prior work has identified a range of security gaps in AI agents. However, it is essential to move beyond reiterating concerns and toward a collaborative, action oriented agenda to mitigate these risks. An international group of leading industrial and academic researchers were gathered by Schmidt Sciences, RAND, and Palo Alto Networks, to contextualize the fragmented cross domain expertise and insights needed to produce solutions that reflect the full landscape interconnected challenges that uniquely arise in the setting of LLM driven AI Agents. This report distills the collective insights from this gathering and contributes: 1) A flexible definition of the functional properties of AI agents, 2) A description of how these AI agent properties create novel implications for security, and 3) An open roadmap to producing interconnected comprehensive solutions.
Title: Context Rules! Privacy and Security for Future Trustworthy AI Agents
Eugene Bagdasarian, Assistant Professor @ University of Massachusetts Amherst, Researcher @ Google
Eugene Bagdasarian is an Assistant Professor at University of Massachusetts Amherst and a Researcher at Google. His work focuses on studying attack vectors in AI systems deployed in real life and proposing new designs that mitigate these attacks. Previously, he received the Distinguished Paper Award at USENIX Security and Apple AI/ML PhD Fellowship.
New AI agents integrate with complex systems and users’ data, thus opening new attack vectors. Worse, security designs struggle with the versatility of agents: booking a trip requires different controls than responding to an email. In this talk, I propose to ground agentic privacy and security in the theory of Contextual Integrity, which defines privacy as appropriate information flows under contextual norms. We use language models to infer the current trusted context and synthesize restrictions on tools and data, then develop a policy engine to deterministically enforce them, helping to isolate attacks that abuse agentic capabilities and data access. While promising, this design raises new questions: from establishing trusted context and improving policy generation to collecting social norms and resolving context ambiguity.
Title: What Unlearning Tells Us About Machine Learning
Katherine Lee, Researcher @ OpenAI
Katherine is a researcher at OpenAI. Her work has provided essential empirical evidence and measurement for grounding discussions around concerns that language models infringe copyright, and about how language models can respect an individuals’ right to privacy and control of their data. Additionally, she has developed large language models (T5), developed methods of reducing memorization, and studied the impact of data curation on model development. Her work has been highly awarded at venues like: NeurIPS, ICML, ICLR, and USENIX.
Abstract To Be Announced