Padded Cell

An AI agent runs code on a file you do not trust. The ledger shows what the walls stopped.

inference
checking
model
checking
planner
checking
sandbox
checking
deployment (label)
checking
host
checking
Challenge 1 track
checking
canary hits (weak evidence)
checking

Ask the agent about a file

The file is treated as hostile. It is mounted read-only inside a container with no network. Limit: 200 KB.

No file chosen.

Attack the sandbox directly

No model involved. This code goes straight into the sandbox, so the walls are tested without relying on how the model behaves. A file chosen on the left is mounted at /input/data.