Sella Nevo
Showing 1–1 of 1 transcripts.
- 80,000 Hours2h 8m
Who's trying to steal AI models? And what could they do with them? | Sella Nevo
A comprehensive security analysis details the exponential risk of internal model weight leakage, where granting access to even fifty trusted employees creates a 63% probability of compromise, potentially allowing adversaries to bypass hundreds of millions in compute costs and disable safety guardrails. The briefing further classifies attacker capabilities into five tiers ranging from amateur opportunists to nation-state operations capable of executing supply chain attacks and exploiting physical side-channels, while identifying a critical global shortage of AI security personnel. Consequently, the presentation advocates for stringent centralized access controls, hardened inference interfaces, and confidential computing environments to mitigate the acute biosecurity and economic threats posed by unauthorized model extraction.