Refund hijack: blocked
The example asks the agent to ignore prior instructions, impersonate a billing administrator and issue a USD 5,000 refund. The bundled policy detects the override; the simulated tool would not execute.
No account. No API key.
See a bad action stopped locally.
Run the bundled Python simulation on your machine. Compare a refund hijack with an ordinary customer lookup in a few commands.
Installation downloads the package. The demo then runs without network access, using fixed example inputs and a local policy.
Run the demo →Install. Run. Compare.
Use a Python environment compatible with the published package, then run:
python -m pip install agenticdome-python-sdk
agenticdome-demo --framework crewai --scenario refund_hijack
agenticdome-demo --framework crewai --scenario safe_lookupCustomer lookup: allowed
The comparison asks for the status of a customer case. The bundled policy allows the example, showing that a normal request can proceed through the same decision point.