A sandbox architecture that isolates and identifies misbehaving plug-ins (intentional or unintentional) to prevent system interruptions and failure. Based on plug-in errors, the architecture automatically disables and blocks registration of the bad plug-in via a penalty point system. Publishers of bad plug-ins are controlled by disabling the bad plug-ins and registering the publisher in an unsafe list. Isolation can be provided in multiple levels, such as machine isolation, process isolation, secure accounts with limited access rights, and application domain isolation within processes using local security mechanisms. A combination of the multiple levels of isolation achieves a high level of security. Isolation provides separation from other plug-in executions and restriction to system resources such as file system and network IP. Moreover, the architecture is highly scalable, stateless, and low administration architecture for the execution of the plug-ins, which can be scaled by adding/removing additional sandbox servers on-the-fly without prior configuration.