There are several common types of problems with challenges; the Sandbox is good at catching some but not others
Things the Sandbox is good at
Where the Sandbox really shines is in circumstances where a challenge breaks a challenge design principle that most experienced PPCG users are aware of. For example, it's fairly easy to go through the recent Sandbox entries and comment on all the ones that require large fixed strings in their output, but are ostensibly about something other than compression. There are a number of similar issues which are almost always quickly caught in the Sandbox.
Another situation in which the Sandbox works well is in cases where the whole concept behind a challenge is poorly conceived. Most such challenges immediately receive negative feedback, and even though a couple of attempts are often made to fix them, it becomes clear that the challenge could never work. As such, these challenges tend not to reach the main site at all, meaning that the Sandbox is doing a good cleanup job in these cases. This might not seem encouraging as a question asker – the only real benefit they get from it is a slight reduction in the (very low on PPCG) chance of a question ban, when the challenge inevitably gets closed – but it certainly helps out the question answerers.
The final way in which I'm aware of the Sandbox helping well is in checking for duplicates. Duplicates are often almost impossible to search for on PPCG, so you typically have to rely on someone who saw the earlier challenge to remember that something was up. (My favourite example of this is this comment, in which a PPCG user remembered competing on an earlier duplicate challenge and already had a solution to it in a file on their own computer, and yet couldn't find the challenge itself. We found it in the end, but it was way harder than it should have been.) It's often possible to get a suspicion that a challenge might be a duplicate, and Sandboxing it with an explicit request for a dupe check often produces useful results.
Things the sandbox is bad at
Now, what the Sandbox is worst at is finding ambiguities in the specification (i.e. the situation for which you're trying to use it); there's definitely a problem here, because this is something that is commonly perceived as something that the Sandbox does/should do. Although I don't know for certain, I suspect there are two reasons for this.
One is that finding an ambiguity often needs a lot of knowledge of specific programming languages; I can see a challenge that says "in situation X, return Boolean false; your program must not error" and immediately realise "there are programming languages in which Boolean false is a type of error", but it's not a fact that the programming community is generally aware of, so most people wouldn't be able to catch that specific issue in the Sandbox. Likewise, there are a large number of potential issues which could only be caught by specific other people. In other words, the issues that I mention that the Sandbox is good at can each be caught by a large proportion of PPCG users; however, ambiguities can often only be noticed by a much smaller subset, meaning that in order to remove all ambiguities from a Sandboxed challenge, you'd probably need around half of PPCG to review the challenge, and even then you'd run the risk of missing some. (Note that posting to main isn't much better at catching this sort of ambiguity, but if nobody catches it on main, the ambiguity turned out not to actually matter.)
The other issue is that some ambiguities only become clear when trying to actually solve the problem, something you aren't really meant to do with Sandboxed problems. (I rarely try to solve even my own problems before they're posted on main, to keep things fair; it's not uncommon for a challenge to change between the sandboxed and main version anyway, so working on an answer in advance is inadvisable as it is. The main exceptions here are solving Sandboxed challenges in Jelly in my head in order to determine if they're too trivial – if I can do it, the challenge probably needs to be harder – and situations in which I write a program I'm particularly proud of and post a challenge to PPCG simply as an excuse to be able to post the program in question as an answer.) Actually writing code can give you a better appreciation for where the corner cases are than looking at the spec.
Ways in which we can improve the situation
So, what can we do about this? First, we probably shouldn't be downvoting challenges unless they're actually downvotable according to SE rules ("this question does not show any research effort; it is unclear or not useful"). Sandboxing questions correctly shows effort into making a good question, and that the question is not immediately unclear; lack of utility (in PPCG terms, being uninteresting to compete on) is also something that you'd hope would at least be commented on in the Sandbox. Incidentally, you can help deal with this issue both as a voter and as a challenge poster; voters can simply not downvote challenges unless the challenge deserves it; and challenge posters can use the trick of linking to the Sandbox post from the main post (even though doing so really shouldn't matter), in order to give a subtle hint of "if you want to complain about this, you should have done so while it was Sandboxed" (which in my experience actually helps to avoid downvotes). For what it's worth, I've Sandboxed a few challenges even though I knew they'd get no useful feedback, purely to be able to link to the Sandbox and say "see, I Sandboxed this!". (Much of the time, though, I am expecting the Sandbox to be able to improve the post.)
That basically just fixes the issue with downvotes, though; it doesn't fix the issues with the amount of exposure that some posts need to resolve ambiguities being lower than the amount that the Sandbox can reasonably give it. I know that I suggested sandboxing posts as on-hold posts on main a while back; the suggestion was generally well-received on average, but controversial. I've since come to view it as a suggestion that would probably be beneficial overall, but which is fairly drastic (and which would likely have significant UI issues), and therefore should be avoided unless necessary; however, if we view the Sandbox as being on the point of collapse, we might have to go through with it. (It's not like the reopen queue is at all struggling at the moment.) That said, there seems to be a rough, but far from unanimous, consensus that the Sandbox is in fact functioning mostly correctly at the moment.