That doesn't mean an employee has access to your account. If you lose your USB dongle, you'll need to prove your ownership of your account, but that doesn't mean G employees will go in and look at your emails.
There are two forms of this assurance.
The first is running into some guy at a party, or on an internet message board, who either works at Google or "knows a guy" who does, who then spends thirty or forty minutes semi-drunkenly rambling at you about how the security at Google is "military grade" and everything is "locked down" and they have some guy who used to run the Mossad in charge of internal security and he's the best guy ever and nothing or no one ever breaks in and that shit's a fortress. Trust the semi drunken ramblings of this guy you just met, your secrets are safe with the big G.
The second form of this assurance is only giving the third party indistinguishable encrypted blobs of data. They can look at them all they want, it doesn't mean anything to them. You don't need to trust anyone, let alone some guy you found on the Internet and transitively a few hundred anonymous computer programmer and IT janitors spread across the world.
Go figure, some people are more persuaded by the second argument. Even if you're reading this and you work at G and you know that the ex-Mossad security guys are digitally looking over your shoulder and you shake your fist at this screen muttering "you don't know how much we have it locked down here, clueless Hacker News commentator" you have to realize, somewhere in your heart of hearts, that a zero trust solution is preferable when you are trying to maximize for security.
Just get a pretty locked down Gmail account and use pen and paper or OTPs or GPG for things that actually need to be secure.
1) Emails disclosed as part of a legal discovery process that compelled Clinton to hand over the emails herself.
2) Copies of her emails acquired from a correspondent with a much-less-secure configuration of their Gmail account than Clinton's configuration of her private server.
Email is, by its nature, a distributed infosec problem.
The threat of a breach of the email account is that it's the key to everything else -- pretty much any other account, with any other service, is instantly accessible the moment you're in someone's email, because the email inbox is where credential-reset requests are sent.
There’s no such thing as zero trust for email unless you’re going to run your own server (even then it’s iffy because of communication with other untrusted servers).
By using a mail service, you have to trust the people running the service. Even if they pinky promise to immediately encrypt every message with your public key a never never never look at it, you cannot know that they aren’t secretly redirecting a copy to the NSA and publishing another copy on the novelty Twitter account @muninmails.
https://moxie.org/blog/lavabit-critique/
Unlike the design of most secure servers, which are ciphertext in and ciphertext out, this is the inverse: plaintext in and plaintext out. The server stores your password for authentication, uses that same password for an encryption key, and promises not to look at either the incoming plaintext, the password itself, or the outgoing plaintext.
The ciphertext, key, and password are all stored on the server using a mechanism that is solely within the server’s control and which the client has no ability to verify. There is no way to ever prove or disprove whether any encryption was ever happening at all, and whether it was or not makes little difference.
... The operator can at any time stop averting their eyes, an attacker who compromises the server can log the password a user transmits, and an attacker who can intercept communication to the server can obtain the password as well as the plaintext email...The cryptography was nothing more than a lot of overhead and some shorthand for a promise not to peek. Even though they advertised that they “can’t” read your email, what they meant was that they would choose not to.
There are maybe a few things that could be done to reduce your trust in "Honest Munin's Very Secure E-Mail Hosting and Used Cars", like, I could publish the code to the server and run it in SGX, allowing anyone in the world to use SGX attestation to verify* that the code running in SGX is the code that I publish, and that code could only initiate encrypted network connections with other mail servers. So even if I ran that in good faith and sat outside the SGX container listening to the network traffic, I'd still be getting a face-full of encrypted data.
This is still a "verify*" because maybe something is funky with SGX. You're not really trusting me now though, you're trusting Intel and the SGX TCB. Maybe that's worse? I don't know. I still think that's an improved proposition over trusting the honesty of system administrators to not grep your INBOX on a whim, though.
edit: I guess this also kind of breaks down because the story for TLS in SMTP land is kind of dire, so you could probably actively MITM the communication between the SGX blob and the outside and there's no practical way for anyone to know that this is going on. Maybe someone will fix that someday! Probably not though!
As for SGX, it doesn’t really solve for this problem. It would allow Google to verify what software is running, but with a Google-controlled network in between the server in question and the user, I’m fairly certain any assurances SGX could yield are gone. Google could publish their code publicly, allow users to request attestation, but send attestation requests to server A running the public code while serving the user’s email on server B running NSA-modified snooping code. And of course your email isn’t actually sitting on a single machine anyway. It’s probably stored redundantly on 3+ backend machines and served by any of thousands of front end machines. Verifying each of these would be infeasible even if technically possible. (I’m not very familiar with SGX, though, so maybe there’s some magic I’m missing here.)
Also attestation is for binaries. I doubt you could actually verify a source match.
There is magic you are missing. I don't want to be the person that glibly says "you should read more about this" and then vanishes, but because I don't have time to write out what the magic is, that's exactly what I'm going to do :(
The magic does allow users to verify that the software they're talking to is what it says. Specifically, it also allows users to set up an encrypted channel with the processor that is running the software that is attested, so you exactly can't do the attack you propose of shuttling attestation to one system and actual data to another.
The technology is pretty cool, you should check it out!
1. Intel does not bill the technology for this purpose, which makes me doubt its fitness. SGX is very specifically billed as being for developers who want to run secure software in an untrusted environment, not for end users who don't trust the developers. These are wildly different scenarios.
2. The "encrypted channel with the processor" implies direct access to individual machines, which is impractical at Google scale.
3. The attestation is for binaries and not source, so attesting that the binaries are unchanged doesn't help users verify that the binaries match the trusted source. (Of course Google also doesn't publish the GMail source anyway.)
Obviously still doesn't help for the GMail case unless Google starts publishing their code, and they pretty much give up on email search, and somehow the whole chain of email cross-provider lands in these enclaves, but it's very interesting.
I guess there's a distinction to be made... perhaps if you go through that process you can regain control of your handle but not private data. But that wouldn't be useful since unless you're using a system which warns that your identity has changed (e.g. a new pgp key) then old contacts will continue to contact you, spilling the beans on old conversations?
in the 80's I had root on all the machines for the main UK ADMD and the level beyond root on the two billing systems - the OS was based on ITS A custom version of PR1MOS which is why there was a level beyond root.
Of course these days I would have had to have SC or DV clearance - I trust those googlers who do have access are properly security cleared :-)