TLDR: it might work, but it will probably struggle, or not work at all, depending on how optimal your "other" policy is

So, strictly theoretically speaking, as long as the behavioral policy explores all the state-actions involved in the optimal policy, DDPG will indeed work (in other words, if the optimal policy involves an action that your safety module excludes, don't expect DDPG to learn it)

Now, we don't live in a ideal world, and we have to face the reality of the approximations made along the way; and practically, it probably won't, or at least, it will struggle a lot

Indeed, DDPG is part of the Q-learning algorithms, which is, by definition, off-policy, which means that you can learn from experience coming from another policy... however, the state distribution of the behavioral will bias the Q estimate, and thus increating instabilities for the learning