Cybersecurity Sandboxing

Sandboxing runs potentially unsafe software inside an isolated, tightly controlled environment rather than allowing unrestricted access to the host system. Its primary security function is containment: it limits the software's permissions and resources so that harmful behaviour can be observed without normally affecting real data, systems or networks.

How sandboxing provides security

A sandbox creates a boundary between the tested software and operational assets. Security controls mediate what the software may read, modify, execute or communicate.

• Isolation separates the test process from the host operating system, applications and sensitive data.
• Least privilege restricts access to files, memory, devices, credentials, networks and application programming interfaces.
• Monitoring records system calls, file changes, network traffic and other behaviour for analysis.
• Disposable environments, such as virtual machines or restricted containers, can be reset after testing.

Use in testing software and AI systems

The tester supplies controlled inputs and observes how the system behaves under normal, adversarial or unexpected conditions. For an AI agent, the sandbox can restrict tool use, code execution and external communication while preserving logs for evaluation.

• It reduces the consequences of malware, faulty code or unsafe autonomous actions during testing.
• It permits controlled experimentation before software receives wider permissions or deployment access.

Limits of protection

Sandboxing reduces risk but does not establish that a system is safe. A sandbox escape may exploit weaknesses in the isolation mechanism, while environment-aware software may conceal malicious behaviour when it detects a test setting.

• The sandbox must avoid exposing real credentials, sensitive data or unrestricted network access.
• Testing cannot cover every input or prove that an AI system is reliable, aligned or free from harmful behaviour.

How UPSC asks this