> VM are a problem waiting to happen.
Something something security threat model and personal risk tolerance...
For a project idea I've been considering how to attach a sensitive (but powerful) machine to the network with the smallest possible interface (in terms of code exposure). RS232 is probably the simplest yet still functional without getting into niche hardware standards.
On balance I think USB might be worse than Ethernet[2] or even Ethernet+TCP/IP just because typical network stacks have probably seen more scrutiny than USB stacks. But at the end of the day I know I'm going to want to use something like seL4 to protect the application from the transport stack(s). And even then you still have to worry about the hardware. There's a real dilemma in choosing the simplest hardware that is still viable in terms of performance, support (including ease of porting or writing drivers), and long-term availability.
[1] Assuming a simple ethernet controller with all the fancy features disabled or, ideally, not even present.
https://arstechnica.com/gaming/2015/01/pokemon-plays-twitch-...
They can contain malicious code just as well as a thumbdrive. The concern with USB is the firmware, and the thumbdrive pretending to be a HID device or having an extra partition with malicious code.
Purchasing many similar drives of a well known brand in a random supermarket is a reasonable alternative.
Just keep them for 1 year before using them, in case an exploit becomes popular - but this is extremely paranoid.
Of course.
> The concern with USB is the firmware, and the thumbdrive pretending to be a HID device or having an extra partition with malicious code.
Right, that's the concern I was alluding to above. You'd still be better off burning optical media (and disabling any autoplay crap on the target system). You can still store severl GB of data on optical media.
> Just keep them for 1 year before using them, in case an exploit becomes popular - but this is extremely paranoid.
This may work if you only stick with very popular makes/models, since they would have a higher chance of being 'noticed' for doing something wrong.
Security is in the spotlight, not in the shadows
Prefer tools that everyone else uses, unless you have a very compelling reason not to (windows: too many 0 days, complexity, and surface of attack ; wifi: too easy to start a rogue network)
Because when these popular tools get hacked, you will know it. I keep an eye on rogue firmware. I have not seen a real world threat yet, but I do expect spear fishing attacks for people involved in crypto (fortunately, not for us just writing software)