How AI Systems Leak the Data They're Built On
Details
Please join Charlottesville Data Science for the talk How AI Systems Leak the Data They’re Built On from Hunter Moore, co-founder and CTO of Hardshell, a Charlottesville-based AI security startup. Hunter’s talk will walk through how deployed AI systems are attacked in practice, such as indirect prompt injection, system prompt extraction, and privacy attacks that induce a model to divulge sensitive data.
We'll be gathering in person at Vault Virginia on the evening of Wednesday, November 11. We look forward to seeing you there!
→ We're moving our events calendar to Luma! We'll continue cross-posting our events here on Meetup for the foreseeable future, but we'll use Luma for planning and all event-related communications. Please head over there to RSVP, and consider subscribing to our Luma calendar! ←
About the talk
Retrieval-augmented generation made it easy to point a language model at an organization's documents. It also made every one of those documents a potential instruction. Sure, the model was trained to refuse harmful requests, but what happens when the attack arrives inside an email, a web page, or a PDF the system retrieves on its own?
This talk walks through how deployed AI systems are attacked in practice – indirect prompt injection, data exfiltration through rendered output, system prompt extraction, and privacy attacks that pull sensitive data back out of a model. Each one is grounded in a real case, from hidden web text that hijacked Bing Chat to the ruling that held Air Canada responsible for what its chatbot told a customer. We will close with which defenses hold up, which ones break under an adaptive attacker, and what a data science team can do about it right away.
No security background required. Familiarity with LLMs and RAG will help.
About the speaker
Dr. Hunter Moore is the founder and CTO of Hardshell. He holds a Ph.D. in Systems Engineering from the University of Virginia, with expertise spanning AI security, machine learning, and autonomous systems.
He has over five years of experience advising the U.S. Department of Defense on AI security, including contributions to foundational policy and protocol development at the Joint AI Center and the Chief Digital and AI Office. His work has encompassed domains such as trustworthy and robust AI, generative AI, predictive and explanatory modeling, reinforcement learning, human systems, and systems engineering of autonomous systems.
He has taught AI-related courses for the U.S. Air Force, U.S. Army Special Operations Command (USASOC), and the University of Virginia, where he also designed and delivers a graduate-level AI security course and supervises capstone teams on trustworthy AI and AI security related topics.
