You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
adcp.signing.verifier.verify_request_signature (adcp 6.6.0, Python 3.12) accepts a request whose Signature-Input header contains the same label twice — e.g. sig1=(...), sig1=(...). The conformance vector negative/021-duplicate-signature-input-label (dist/compliance/3.1.13/test-vectors/request-signing/) expects rejection with request_signature_header_malformed.
RFC 8941 §3.2 permits parsers to reject duplicate dictionary keys, and the vector's $comment cites the AdCP profile's downgrade-protection rule requiring it: with last-wins parsing, two different signature bases can hide behind one label — the verifier validates one of them while a downstream consumer may read the other.
We found this while grading our seller against the vectors: the live failure initially surfaced as request_signature_window_invalid (the vector's frozen timestamps fail the window check first), which masked the parser leniency. The repro below uses a fresh window and a signature that genuinely verifies — the duplicate label is the only defect, and the verifier accepts it.
Likely the same parser layer affects negative/022-multi-valued-content-type and negative/023-multi-valued-content-digest (we haven't isolated those to the parser as cleanly; the duplicate-label case is proven).
SignatureVerificationError with code request_signature_header_malformed, per the vector and the profile's step-6 strict-parse requirement.
We've added an application-side pre-check as defense in depth, so this isn't blocking us — filing so the SDK's parser matches the conformance vectors it ships with. Happy to contribute the fix if useful.
ad-tech-protocol-expert: PR fix(signing): reject malformed structured-field input at step 1 #993's paren/quote-aware pre-parse rejection at step 1 is architecturally correct and non-breaking; KNOWN_VERIFIER_GAPS is empty and the conformance vector is active and unexempted in 7.0.0rc2.
security-reviewer: Comma-joined duplicate-label check covers vectors 021–023; residual WSGI/Flask gap (proxy last-wins dedup before PEP 3333 folding — SDK cannot catch this) is pre-existing and documented in _header_precheck.py, no escalation warranted.
My take: Thanks for filing — solid repro and correct threat-model analysis. This bug is confirmed in 6.6.0 and was fixed by PR #993 (fix(signing): reject malformed structured-field input at step 1), which ships in 7.0.0rc2 and the forthcoming 7.0.0 stable. The conformance vector negative/021-duplicate-signature-input-label is now active in the suite with no exemptions. If you're running Flask/WSGI behind a proxy that does last-wins (not comma-join) deduplication on repeated header names, the SDK module docstring notes that form of the attack cannot be caught at the SDK layer and a proxy-level rule is the recommended mitigation.
Fixed by #993 and shipped in v7.0.0-rc.2: duplicate structured-field labels and repeated singleton headers are rejected during the step-1 raw-header precheck. The conformance vector passes in the full v7 release suite.
Activity
bokelley commentedon Aug 14, 2026
Triage
Classification: Bug
Bucket(s): signing
Status: deferred
Milestone: omitted — fix already in 7.0.0rc2
What the experts said:
KNOWN_VERIFIER_GAPSis empty and the conformance vector is active and unexempted in 7.0.0rc2._header_precheck.py, no escalation warranted.My take: Thanks for filing — solid repro and correct threat-model analysis. This bug is confirmed in 6.6.0 and was fixed by PR #993 (
fix(signing): reject malformed structured-field input at step 1), which ships in 7.0.0rc2 and the forthcoming 7.0.0 stable. The conformance vectornegative/021-duplicate-signature-input-labelis now active in the suite with no exemptions. If you're running Flask/WSGI behind a proxy that does last-wins (not comma-join) deduplication on repeated header names, the SDK module docstring notes that form of the attack cannot be caught at the SDK layer and a proxy-level rule is the recommended mitigation.Triaged by Claude Code. Session: https://claude.ai/code/cse_013CSwE6ovNuL1yty2zyvjRw
Generated by Claude Code
bokelley commentedon Aug 16, 2026
Fixed by #993 and shipped in v7.0.0-rc.2: duplicate structured-field labels and repeated singleton headers are rejected during the step-1 raw-header precheck. The conformance vector passes in the full v7 release suite.