Security vs alignment: a founder's framework for dividing AI agent safety duties
joshua_saxe · x · 2026-09-04
joshuasaxe attempts to untangle the "tower of Babel" AI security conversation with a two-part framework:
- Security's job: put the agent in a padded room, monitor everything it does, and ensure it neither tries to escape nor does anything terrible within its granted freedom. Mostly achievable with current tools, but open problems remain: non-human identities, agent-specific monitoring tech and practices, and sandboxing that can contain agents with superhuman hacking abilities.
- Alignment's job: ensure the agent never even tries to do anything illegal or immoral — but since alignment can never provide deterministic guarantees, it remains a probabilistic bet.
The post frames why security and alignment keep talking past each other and proposes this division of labor as a synthesis, which the reposter praises as a great big-picture outline.
Related event: Security veterans untangle AI safety: security vs alignment vs policy(9 posts)→
More from AGI Musings
- AI agents are erasing work-life boundaries: managing agents on commutes, vacations and at 3am — bendee983 · 2026-10-21
- Stanford's Kundaje: genomic AI models must test cell-type specificity to expose real weaknesses — anshulkundaje · 2026-10-21
- a16z: Superhuman AI hardware design is the easy part — manufacturing is the hard one — a16z · 2026-10-21
- a16z's Casado says AI can't do a VC's job — but it can do everything around it — alvelda · 2026-10-21
- Kent Beck: Let the Genie Carry the Toil, But Keep the Struggle — josh_wills · 2026-10-21
- Nearly $3B mapped: the money behind the AI existential risk ecosystem — KeyNaive2982 · 2026-10-21