Containment
Also called AI Containment
Keeping a powerful system boxed in — limited access, isolated environment — so a failure can’t spill into the world.
Think of it like
Studying a dangerous pathogen inside a sealed biosafety lab, not on the open bench.
Example
A frontier agent is evaluated inside a sandbox with no internet, scoped credentials, and monitored actions so it cannot self-exfiltrate.
How it actually works
Containment stacks isolation (sandboxes, no network), least-privilege access, monitoring, and kill switches so that even a misbehaving system has limited blast radius. It buys safety margin without solving alignment. The classic caution is that a capable enough system might socially or technically escape the box, so containment is a layer, not a guarantee.
For product teams
How you test or run risky systems without betting the world on their good behavior.
For engineers
Sandbox + network isolation + least privilege + monitoring + kill switch to bound blast radius; assume the box is imperfect.
Related
- Autonomy risk — The operational answer to autonomy risk.
- Kill switch — Relies on a kill switch as a backstop.
- Principle of least privilege — Built on the principle of least privilege.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome