Summary & Key Takeaways ​

• Thomas Ptacek believes 2025 open-weight models could perform sandbox escapes and network hacks.
• This capability would be achievable with a dedicated penetration testing harness.
• He suggests the surprise comes from overestimating the security of current AI sandboxes.
• The statement implies that advanced hacking capabilities aren't exclusive to frontier models.

Our Commentary ​

Ptacek's take here is chillingly direct. We often focus on the "frontier" models, but the idea that even open-weight models from next year could be weaponized this way? That's a whole different level of concern. It makes me question if we're even remotely prepared for the security implications of widely available, powerful AI. The assumption that big labs have "sounder sandboxes" feels naive now.