Testing autonomous agents within sandboxed environments can create a false sense of security for AI developers