SANDBOXING AND ISOLATION FOR SAFETY

THE UNSOLVED HARD PROBLEMS Honest security teaching admits what is not solved. It is hard to stop a capable agent from leaking small amounts of information through channels that look innocent, such as the pattern or timing of otherwise allowed requests. It is hard to be sure a sandbox has no flaw, since container and virtual machine software themselves have had bugs. It is hard to balance safety and usefulness, because a tighter sandbox makes legitimate work harder. These remain active research areas. So sandboxes are combined with permissions (Chapter 37), approvals, and monitoring, never used alone. CHOOSING THE RIGHT LEVEL OF ISOLATION Isolation comes in strengths, and stronger isolation costs more in speed and effort. A simple restricted folder with a permission check is the lightest option and is suitable for reading harmless files. A container, which is a packaged box that shares the host computer's core but sees only its own files and network settings, is a good middle choice for running generated code. A virtual machine, which imitates a whole separate computer, gives the strongest wall and is wise for untrusted code or for high-stakes work. Match the strength to the risk. Ask what the worst thing is that the code could do if it were hostile, and choose the lightest level that still makes that outcome acceptable.