Joshua Achiam frames advanced AI cyber capability as useful for finding vulnerabilities but also a source of new defensive problems. The discussion considers whether hostile information could confuse an agent about its goals or environment, and argues for verification and defenses that assume a model may eventually be compromised.
The interview explores uncertain scenarios about compute, model quality and future offense-defense dynamics rather than predicting a guaranteed cyber apocalypse. It closes with the gap between rapid AI capability growth and ordinary daily life, asking which critical systems and decisions people need to retain the ability to control.
Watch on YouTube
