Q: WHAT VOLTAGE IS SAFE?
A: The one the CPU picks on its own. The firmware is actively monitoring and managing more chip health data than you can possibly imagine or digest--every second--to ensure performance is being reliably maximized for n cores under load at y temperature for x time. Stepping in to fix a static vcore, because you "feel" something about what the CPU is doing on its own, or you misread what someone said on the internet, won't help you. It will likely subtract performance (reduced Vmax = reduced perf), and it will likely raise the average voltage the chip experiences over a 24H period vs. what it will self-manage to on its own accord (you've disabled idle voltages). If the purpose of setting a manual vcore was to reduce voltage, you missed: you actually just raised the average voltage.

..I think there's also a credible basis to argue that it's very difficult for monitoring utilities to totally keep up with what a modern CPU is doing. You can do various amounts of averaging, abstracting, and approximating to give a general overview of what's going on, but there are so many sensors updating so fast that it's an absolute firehose of information to take in a total view of everything that's happening.
For example, just looking at temperature, a Ryzen 9 3900X has 20+ temperature sensors that I am aware of. Each core has its own, the caches have their own, the PHYs have their own, and more. All of these are changing in 1ms intervals, which is fine for the CPU's firmware to keep up with, but what does a monitoring utility with its one "CPU temperature" box report? Okay, maybe it takes the warmest temperature of any sensor. Is that "accurate" if it's the outlier? What if that peak only lasted for 2ms on the edge of its polling window? What if just one core is 10C hotter than the rest, should the average be presented? Rolling average? Peak? Instantaneous?

There are lots of ways to slice it, which is why different tools tend to report different values, as each tool developer (including us) has their own position on what is "realistic" versus the dozens of sensors the processor actually has. Now you can apply that same thinking to voltages, frequencies, and whatever else is dynamic.