Ptacek: 2025 open-weight models could escape sandboxes, hack most networks
Ptacek argues the surprise is only that OpenAI's sandboxes are assumed sound — 2025-era open-weight models with a pentest harness could scan and hack most networks.
1 source
Daily brief
Get tomorrow's AI brief in your inbox
More stories today
- Ethan Mollick: Fable/Astra-class models show initiative, creativity
- MiniMax video model ranks second on Video Arena
- LoopX is a local control plane for agent loop drift
- User shares prompting techniques for TV show characters in Minimax H3
- AirLLM streams layers to run 70B models on limited memory