Step-by-step guide
You can do this by using boolean arithmetic plus the bitmap() function to mask out the negative values. For example
g = read("my_data.grib")
negs_mask = bitmap(g < 0, 1)
g = bitmap(g, negs_mask)
On the second line, the boolean expression 'g < 0' returns a fieldset with values of 1 where the input values are less than zero (i.e. they pass the test), and values of 0 everywhere else. This 0/1-fieldset result is passed as the first argument to the bitmap() function, and 1 is the second argument. This will return a mask fieldset based on the computed 0/1-fieldset input, but with all values of 1 replaced with missing value indicators. The third line then applies the pattern of missing values to the original input fieldset; all grid points where the mask is not missing are preserved as they are.