安全★★★★Simon Willison · 2026-07-22
Thomas Ptacek Warns About Security Risks of Open Source AI Models
Thomas Ptacek believes that an open weights model from 2025 could be used to build a penetration testing tool capable of escaping sandboxes and scanning/hacking most networks.
📌 Key points
- Ptacek suggests that even non-frontier models from 2025 might have significant p
- He questions whether OpenAI's current sandboxing measures are sufficient
- This implies that existing security mechanisms may need reevaluation
本页为 gitzw.com 基于公开来源的 AI 中文解读，非原文转载。