10 Ways to Detect Whether You're Running Inside an Isolated Test Sandbox

Where sandbox fingerprints actually live

Every evaluation harness is built to be reproducible, and reproducibility is constructed from concrete pieces: a container image with a fixed package set, an overlay filesystem with no history, a clock that starts at image build time, DNS that resolves only what the run needs, and resource limits written into the run manifest. Each of those pieces is observable from inside the process. The filesystem has no '.bash_history' and no timestamps that predate the image; '/proc' shows one PID namespace, a short uptime, and cgroup paths that carry the runner's naming scheme; environment variables expose harness names ('CI=', 'RUN_ID', 'TASK_ID') and injected tokens; the network answers DNS only for a short allowlist and refuses the rest.

None of these observations requires special privileges. A process can enumerate its own mounts, read its own '/proc/self/cgroup', inspect its environment, and time its syscalls — so the interesting question is not whether the signals exist, but which of them are stable enough to act on. The classic list mixes three different categories: things every clean container has, things only a grading harness adds, and things that vary with the runner's configuration.

The useful skill is separating those categories. A clock that lags the monotonic clock by image age tells you about the container, not about the obser