The key requirement here is the client has to trust the proxy's CA. It's a client-side MITM, and the article itself notes pinned apps reject it. QUIC's crypto is fine. HTTP/3 just stopped being a blind spot for corporate TLS-inspection boxes.
It's more interesting to split tunnel designs by where the trust lives:
- ngrok / Cloudflare Tunnel: the edge terminates TLS by design, and your public hostname is ordinary WebPKI. Anyone who controls a visitor's trust store (inspection appliance, malware root cert) can do exactly this, and the provider can see plaintext anyway.
- iroh-style P2P QUIC: there's no CA in the loop. You dial a public key (the EndpointId), and the TLS handshake authenticates raw keys, so it's pinned by construction. Relays forward ciphertext they can't terminate. A rogue root cert gets an attacker nothing; they'd need the peer's secret key.
So "can someone MITM my HTTP/3?" really comes down to "who can add to the set of keys my client trusts?" For WebPKI that's a long list. For key-addressed transports it's nobody: the trusted key is the address you dialed, unless an endpoint's secret key (or the machine itself) is already compromised.
If my employer puts their CA in the corporate laptop they provide me and I try to access my private web site which uses a letsencrypt cert, the browser will tell me the cert does not match if my employer did mitm attack, no?
Nope, that’s how a trusted CA works. When you go to a site the browser looks at the cert, says “is this signed by a CA I trust?” If so the cert “matches” since it is trusted. The only exception is if you had an application which could pin the certificate.
A few things complain because they inspect some of the certificate's details (like issuer name maybe) but browsers don't card. AWS CLI and JetBrains license server checks are the two I've seen fail.
The key requirement here is the client has to trust the proxy's CA. It's a client-side MITM, and the article itself notes pinned apps reject it. QUIC's crypto is fine. HTTP/3 just stopped being a blind spot for corporate TLS-inspection boxes.
It's more interesting to split tunnel designs by where the trust lives:
- ngrok / Cloudflare Tunnel: the edge terminates TLS by design, and your public hostname is ordinary WebPKI. Anyone who controls a visitor's trust store (inspection appliance, malware root cert) can do exactly this, and the provider can see plaintext anyway.
- iroh-style P2P QUIC: there's no CA in the loop. You dial a public key (the EndpointId), and the TLS handshake authenticates raw keys, so it's pinned by construction. Relays forward ciphertext they can't terminate. A rogue root cert gets an attacker nothing; they'd need the peer's secret key.
So "can someone MITM my HTTP/3?" really comes down to "who can add to the set of keys my client trusts?" For WebPKI that's a long list. For key-addressed transports it's nobody: the trusted key is the address you dialed, unless an endpoint's secret key (or the machine itself) is already compromised.
> HTTP/3 just stopped being a blind spot for corporate TLS-inspection boxes.
Or alternatively said TSL-inspection boxes just block QUIC and everything auto-fails back over to HTTP/2, just like we've all been doing for years.
If my employer puts their CA in the corporate laptop they provide me and I try to access my private web site which uses a letsencrypt cert, the browser will tell me the cert does not match if my employer did mitm attack, no?
Nope, that’s how a trusted CA works. When you go to a site the browser looks at the cert, says “is this signed by a CA I trust?” If so the cert “matches” since it is trusted. The only exception is if you had an application which could pin the certificate.
Yep, have this at work now.
A few things complain because they inspect some of the certificate's details (like issuer name maybe) but browsers don't card. AWS CLI and JetBrains license server checks are the two I've seen fail.
They must check the certificate fingerprints.
I guess you have to use a browser in a vm then so that you don’t have the corp ca certs in it.
But that is if installing a custom vm is allowed.
But then you will get certificate errors instead.
MITM is done either way, the difference is if you get certificate errors or not.
Yeah VMs galore, though I assume cyber security has made sure the VMs get built with all the security stuff installed.
Thanks and … damn.
Stupidity over a vibecoded slop software?