"They" will just spray the machines with liquid nitrogen, pull them out of the rack, put the DRAM in a thermos w/ LN2 and read the data at their leisure.
With modern encryption protocols, this yields you nothing.
The feature is called Perfect Forward Secrecy, and protects past flows from later key compromise.
Wireguard supports this, which is what Mullvad uses. (For some reason, speculation about which is an exercise left to the reader, WPA in Wi-Fi still does not.)
Not exactly nothing, just not ongoing compromise. TLS session keys can be pretty long-lived; I don’t know how long-lived Wireguard’s equivalent keys are, but even a relatively conservative few minutes can yield valuable traffic and metadata.
(That being said, I think having your RAM frozen to extract ephemeral secrets is firmly in the “fully hosed” threat model, and is not a realistic model for 99.9% of users to plan for.)
AMD processors support encrypted RAM, called SME[1]. The key is internal to the CPU and randomized at boot. Sniffing a live RAM chip or reading a perfectly preserved frozen RAM will give you nothing. It's a big part of why the xbox one was never hacked.
You can enable SME in the BIOS on all AMD-based business laptops and AMD EPYC servers.
The word "just" is doing some heavy lifting here... To "just" do this, the agent would need to more or less completely take over the building infrastructure before Mullvad could react which is a lot easier said than done. Even if it were trivial it's still quite a few cuts above any competing VPN service.
https://ieeexplore.ieee.org/document/8388826