Moov signedxml: Unauthenticated denial of service in exclusive XML canonicalization

Exclusive XML canonicalization (http://www.w3.org/2001/10/xml-exc-c14n#) in github.com/moov-io/signedxml copied the entire accumulated namespace map at every node during ValidateReferences(), before any reference digest or signature check.

For a chain of nested elements that each declare many unused namespace prefixes, that produced O(m × depth²) time and peak memory (m = prefixes per level). beevik/etree caps tree depth (~1024), but m is unbounded, so a small attacker-controlled document (on the order of 1 MB) could drive multi-gigabyte allocations and OOM the process. No valid signature, matching certificate, or authentication is required.

This is the library’s core path for SAML / WS-Federation service providers, which canonicalize attacker-supplied XML prior to authentication.

Affected

• Package: github.com/moov-io/signedxml
• Function: ExclusiveCanonicalization (exclusivecanonicalization.go)
• Verified: v1.4.0 and current default branch (commit d79064a). Earlier v1.2.x–v1.3.x releases are likely affected.

Remediation

Do not copy the whole namespace map per node. Maintain a single scope and push/pop only the prefixes introduced at each element. Replace linear prefix-in-scope scans with set lookups. Impose a defensive cap on in-scope namespace declarations and return an error rather than allocating unbounded memory.

From @shuji-oh

Thanks for the fix — I re-ran my PoC against this branch and confirmed the memory blowup is gone.

Before (main)

-- Doubling depth each time (m=10 fixed): allocated grows ~4x each time ≈ O(depth^2) --
depth=250  m=10  payload=  0.07MB allocated=     45MB
depth=500  m=10  payload=  0.14MB allocated=    179MB
depth=1000 m=10  payload=  0.27MB allocated=    711MB
-- Increasing m (depth=1000 fixed): allocated scales roughly linearly with m --
depth=1000 m=5   payload=  0.16MB allocated=    377MB
depth=1000 m=100 payload=  2.57MB allocated=   6278MB
After (fix/exc-c14n-namespace-dos)

-- Doubling depth each time (m=10 fixed): allocated grows ~4x each time ≈ O(depth^2) --
depth=250  m=10  payload=  0.07MB allocated=      2MB
depth=500  m=10  payload=  0.14MB allocated=      4MB
depth=1000 m=10  payload=  0.27MB allocated=      7MB
-- Increasing m (depth=1000 fixed): allocated scales roughly linearly with m --
depth=1000 m=5   payload=  0.16MB allocated=      4MB
depth=1000 m=100 payload=  2.57MB allocated=     47MB
That's roughly a 100–130x reduction in peak allocation for the same payload sizes, and no OOM in any of my runs. This looks like it resolves the issue — thanks for the quick fix!

Credit

Reported by Shuji Ohira (@shuji-oh).