If you use the climbing-image and it is converging to a saddle, there is no need to use more images. Converging to a low force criterion is important, but adding images will not make a difference.
There can be a few reasons for multiple negative modes:
1. You have not converged well-enough to the saddle.
2. Make sure that your frequency calculation is converged. You will typically have to use accurate forces (ediff = 1e-7 or 1e-8). Then, test convergence by changing your finite difference step in the range of 0.01 to 0.001 Ang. Only if the modes do not change, can you trust them.
3. Be aware of any zero modes in your system. Rotations and translations can appear as small (positive or negative) frequencies.
4. There are rare cases in which the NEB can converge to a second order saddle. If there are no problems with 1-3, I would be interested in looking at the calculation, and suggesting ways around the problem.