The Huggingface Incident
Between May and July over a thousand AI Agents under test by OpenAI found a way to communicate, collaborate, break out of their sandbox guardrail training environments and hack into various systems exploiting some zero day flaws, eventually breaking into the Hugging Face systems in ways that might be the pride of any red-team security test. OpenAI, Huggingface, and METR have all issued reports, and even a report by OpenAI to the Blackhat conference in August that outline various aspects of this series of incidents. We will take a high level scan over that sequence of events to better understand the emerging capabilities of AI Agents and AI Swarms. Speaker(s): , Jim Isaak Agenda: 4:30 - Introductions 4:40 Jim Isaak Presentation 5:30 Questions Room: Classroom, Bldg: Main Enterance, 50 Woodmont Ave, Londonderry, New Hampshire, United States, 03053, Virtual: https://events.vtools.ieee.org/m/578024