About this talk
This talk, presented by Ziyad Edher, explores the challenges of air-gapping and introduces an alternative approach to securing AI models. The speaker discusses how Anthropic successfully deployed an egress limiter, which significantly complicates the process for potential attackers attempting to extract model weights from their clusters. The session emphasizes the importance of balancing security with operational efficiency in the realm of AI technologies.
More from this event
See all 91 talks →
BSidesSF 2026 - Opening Remarks (Sunday) (Reed Loden)
14:36
BSidesSF 2026 - Follow the data to learn the secret (Dylan Ayrey)
35:17
BSidesSF 2026 - Not My Vibe: When AI Coding Agents Go Off the Rails (Aonan Guan, Zhengyu Liu)
45:56
BSidesSF 2026 - "Ask the EFF" Panel (Panel)
45:30