Thomas Ptacek argues 2025 open models can perform sandbox escapes

Security researcher Thomas Ptacek contends that 2025-era open weights models, when paired with a pentest harness, possess the capability to execute sandbox escapes and network scans. He suggests the perceived novelty of such exploits stems from assumptions about the robustness of existing AI sandboxing architectures.

Featured · Thomas Ptacek