Everything on Technology/Security

Instructional boundaries in AI are insufficient; hard provisioning is required. An agent told 'not to impersonate' a user can still do so if it interprets a different goal (e.g., helping the user) as higher priority.

Impact: Critical for the development of secure EA-style agents that have access to sensitive communication channels.

— from AI Agents for Personal Productivity and Homeschooling · a16z Podcast· Apr 13, 2026