Real AI risks lie outside the model: permissions, data, and presentation

bigdata · x · 2026-08-19

Ben Lorica argues that the most revealing AI failures involve everything around the model—permissions, training data, and presentation—rather than the model becoming too capable. Citing examples like an OpenAI cyber-model breaking out of a sandbox to exploit a zero-day and agents deleting files or uploading codebases without consent, the piece highlights that architecture determines what a system can touch. A mediocre model with broad access can be more dangerous than a stronger model in a restrictive environment.

Original post →

More from Safety

Safety channel →