What I found
MsQuic accepts a PATH_RESPONSE with a raw byte compare. A newly created path has zeroed challenge storage. That means an authenticated peer can send an all-zero PATH_RESPONSE from a claimed migration tuple before the stack has ever generated a PATH_CHALLENGE.
The path is marked peer-validated. Anti-amplification allowance jumps from a few dozen bytes to UINT32_MAX. RFC 9000 requires the response to echo data from a challenge that was actually sent. This matcher never checks that a challenge exists.
sequenceDiagram
participant Peer as Authenticated peer
participant MsQuic
participant Host as Claimed address
Peer->>MsQuic: Ordinary handshake, 1-RTT keys
Note over MsQuic: New path Challenge is still 0x00
Peer->>MsQuic: 1-RTT from new tuple plus all-zero PATH_RESPONSE
MsQuic->>MsQuic: memcmp matches zeros, path marked valid
MsQuic-->>Host: Application data without anti-amp bound
How the state machine fails
A 1-RTT packet from a new source tuple goes through QuicConnRecvDatagrams → QuicConnGetPathForPacket. QuicPathInitialize leaves Challenge, SendChallenge, and PathValidationStartTime at zero.
Authenticated frame processing then does:
if (!TempPath->IsPeerValidated &&
!memcmp(Frame.Data, TempPath->Challenge, sizeof(Frame.Data))) {
QuicPathSetValid(Connection, TempPath, QUIC_PATH_VALID_PATH_RESPONSE);
}
There is no “validation has started” predicate. PathValidationStartTime != 0 is already used later, when the challenge is actually scheduled. Eight zero bytes compare equal to eight zero bytes, so QuicPathSetValid fires and the path allowance becomes UINT32_MAX.
flowchart TD
Pkt["1-RTT from a new tuple"] --> Recv["QuicConnRecvDatagrams"]
Recv --> Path["QuicConnGetPathForPacket"]
Path --> Init["New QUIC_PATH Challenge = 0"]
Init --> Frames["QuicConnRecvFrames"]
Frames --> Cmp{"memcmp response vs Challenge"}
Cmp -->|"all-zero equals zeros"| Valid["QuicPathSetValid"]
Cmp -->|"wrong bytes"| Keep["Path stays unvalidated"]
Valid --> Amp["Allowance = UINT32_MAX"]
Attack shape
- Complete a normal QUIC handshake and hold 1-RTT keys.
- Send a fresh protected packet from a claimed new address and port.
- Include an all-zero
PATH_RESPONSEbefore the server starts validation for that tuple. - Optionally include a non-probing frame such as
PINGso the now-validated path becomes the migration target.
No operator privilege or unsafe configuration is required. The attacker must already be an authenticated peer and must be able to spoof a source tuple. Those bounds limit volume; they do not restore return-routability.
Impact
The protected control is the new-path send limit:
min(queued application bytes, congestion/loss limits, 3 × bytes received on the unverified address)
After the premature validation it becomes:
min(queued application bytes, congestion/loss limits, UINT32_MAX)
Target-native accounting showed validated=1 allowance=4294967295 for the all-zero pre-challenge response, versus validated=0 allowance=27 for the wrong-response control. Ten 1,200-byte send records were accepted against a secure 27-byte maximum.
This is a third-party reflection / DoS primitive, not a crash and not RCE. Actual reflected volume still depends on queued application data, congestion control, MTU, and spoofing — none of which replaces proof that the peer owns the claimed address.
Fix
Reject a PATH_RESPONSE unless validation is outstanding. The existing lifecycle flag is enough:
if (!TempPath->IsPeerValidated &&
TempPath->PathValidationStartTime != 0 &&
!memcmp(Frame.Data, TempPath->Challenge, sizeof(Frame.Data))) {
QuicPathSetValid(Connection, TempPath, QUIC_PATH_VALID_PATH_RESPONSE);
}
That preserves a legitimate in-flight challenge/response and closes the zero-initialized window.
What this is not
Not 0-RTT. Not a TLS-provider bug. Not a design choice to trust authenticated peers with unvalidated addresses — RFC 9000 §8.2.3 and §9.3.1 still require the challenge to have been sent. Congestion control is not a substitute for path validation.