Thanks to the features recently added in HWInfo, it is rather easy to verify what kind of voltages are safe to your CPU specimen. As said before, each and every CPU is different in terms of silicon characteristics, even if they are the same SKU with consecutive serial numbers.
First, set PPT / TDC and EDC to sufficiently high values, which are unreachable in practice. Then, make sure that the CPU is running bone stock (outside the altered PBO limits). Practically meaning that there are no fixed frequencies, voltages, voltage offsets or load-line adjustments used.
Open HWInfo, go to the "Central Processor(s)" and make sure that "CPU PBO Scalar (Reliability Reduction)" reads 1.00x. Then run the worst-case multithreaded workload of your choice (the worst-case workload of your use). While the workload is running, check HWInfo sensors for the "CPU Core Voltage (SVI2 TFN)". That value is the practical one, which the silicon fitness monitoring has allowed and is safe, without loosing any reliability. You can double check that this value wasn't affected by any of the other limits (thermal, power, current): change PBO Scalar to 2x value and repeat the test. If the observed voltage has increased from doing that, then the figure is accurate for the workload you used.