These distributed randomness beacon servers can also provide private random data. There exists a project with daemon  which collects entropy from distinct sources and produces secure output, which can be used to reseed Linux /dev/random and /dev/urandom. It may be useful for devices without reliable HW RNG.
 - https://www.cloudflare.com/leagueofentropy/
 - https://github.com/Snawoot/drb-client
is not exactly what you're looking for out of the box, but it's open-source and it includes a HWRNG, so you could make it do what you want with some custom firmware. I would be happy to discuss a custom development for you if you don't want to do it yourself.
Disclosure: the SC4-HSM is my product.
Calling a device an HSM comes with an expectation that it will resist tampering or intensive analysis. A standard microcontroller on an exposed PCB cannot fulfill that expectation.
The STM32F405 is a standard microcontroller. Its security features are generally comparable to those which are present on other STM32 microcontrollers, as well as most microcontrollers in general. The PCROP (proprietary code readout protection) features which you appear to be referencing here don't help, as your application code is already open-source, and the key data that you're trying to protect must be readable to be used. Moreover, PCROP functionality is absent on the STM32F405.
STM32 readout protection is not particularly robust. There are a number of known attacks which can be used to bypass it on certain families without decapping the chip. I'm not sure if there are any attacks currently known against the F4 series, but I wouldn't doubt that they exist.
Most STM32 parts actually have a tamper detection feature which can be used to clear the contents of a battery-backed RAM when triggered. This sort of configuration still doesn't make a HSM, but it would be a step in the right direction.
The SC4-HSM costs $75. It is intended to compete against the likes of the Trezor, the Ledger, and the Yubikey, not the SafeNet Luna. Its primary threat model is a compromised host, not loss of physical possession. Is it the most secure HSM you can buy? No. Is it the best choice for use in a data center? No. Is it secure against loss of physical possession to a state actor? Probably not (though you can always choose to encrypt the keys, which has a pretty good shot of holding up even against state actors). Is it fair to say flat-out, "It's not an HSM" implying that it provides no security at all? No.
If the bus snooping attacker can break the encryption he could do so with the over the wire copy too.
Is it a question of _authenticating_ the data from the device? If so-- avoiding the case where the bus attacker would replay randomness would require great care in protocol design.