feat(aws): sign SigV4 requests through the system OpenSSL by szedan-rh · Pull Request #1351 · praxis-proxy/ai

szedan-rh · 2026-09-24T14:56:44Z

Summary

aws_sigv4_sign now computes its signature through the system OpenSSL
(EVP digest and signing APIs), so it can ship in the FIPS build. The
aws-sigv4 crate hard-wires RustCrypto hmac/sha2 with no backend
seam and keeps its canonical-request types private, so the SigV4
protocol steps move into filters/src/aws/signing.rs and the crate
becomes a dev-dependency used as the test oracle.

praxis_ai_apis::hash::HmacSha256 next to the existing OpenSSL Sha256
aws-sigv4-filter added to FIPS_FEATURES and Containerfile.fips
make test-fips-provider: unit tests with the RHEL FIPS provider active
docs updated (docs/fips.md, docs/features.md, FIPS tooling)

No change to filter configuration, headers set, or error handling.

Related issue

Refs #814

Validation

24 aws:: unit tests, including a 14-case differential test that
compares every emitted header byte-for-byte with aws-sigv4, the AWS
published IAM key-derivation vector and the S3 example
Same tests with the RHEL FIPS provider active in the test process
make lint (includes fips-deps: no denied crate with the filter enabled)
Integration example test aws_sigv4
Default graph shrinks from 426 to 423 crates

make test-fips still fails the 14 pre-existing agentic_loop tests, which
fail on main as well (see #1335).

Breaking changes

None.

aws_sigv4_sign computed its HMAC-SHA256 through the aws-sigv4 crate, which hard-wires the pure-Rust hmac/sha2 crates and offers no way to substitute them, so the filter had to be left out of the FIPS build. Move the SigV4 canonicalization and key derivation in-tree and compute SHA-256 and HMAC-SHA256 through the OpenSSL EVP seam in praxis_ai_apis::hash. aws-sigv4 becomes a dev-dependency: a differential test checks every header the signer emits against it, alongside the AWS published vectors. Add aws-sigv4-filter to the FIPS feature set and a test-fips-provider target that runs the unit tests with the RHEL FIPS provider active. Filter configuration and behaviour are unchanged. Refs praxis-proxy#814 Signed-off-by: szedan <szedan@redhat.com>

leseb

P1 — Restore Amazon’s copyright attribution. signing.rs ports substantial aws-sigv4 code nearly line-for-line but replaces its Amazon copyright notice. Apache 2.0 §4 requires retaining applicable notices in derivative source. Keep the Amazon notice and identify Praxis’s modifications before merging. Apache License §4

aslakknutsen · 2026-09-24T15:55:14Z

Update the doc?

Yes, missed that one. Dropped SigV4 from the list in 09a2745.

…comment Signed-off-by: szedan <szedan@redhat.com>

szedan-rh · 2026-09-24T16:08:53Z

Fair point on the attribution. 09a2745 restores the Amazon notice in signing.rs and states what was ported (path normalization, encoding set, excluded headers, canonical request assembly from aws-sigv4 1.5.3 / aws-smithy-http 0.64.0) and what changed (the crypto now goes through the system OpenSSL).

alexsnaps

aws-sigv4 (+ transitive hmac/sha2) 🔥 Good stuff!

alexsnaps · 2026-09-24T16:21:58Z

+    /// duplicate and excluded headers, spaces, query ordering, ports,
+    /// tokens, regions, services and dates).
+    #[expect(clippy::too_many_lines, reason = "test data: one entry per canonicalization edge")]
+    fn differential_cases() -> Vec<Oracle<'static>> {


Might want to eventually harden this around canonical_query edges maybe, but I think that's not a real problem right now.

alexsnaps · 2026-09-24T16:24:36Z

+        if written != SHA256_LEN {
+            // Cannot happen for SHA-256; treat it as the library misbehaving
+            // rather than returning a truncated tag.
+            return Err(ErrorStack::get());


This is purely hypothetical as of now... wondering how much a static message would read better than an empty stack. But, again, purely hypothetical (

The test-fips-provider target passed OPENSSL_CONF through cargo's runner, which OpenSSL ignores when the path is wrong, so the tests could run on the default provider and still pass. The run now also sets PRAXIS_TEST_FIPS_PROVIDER, and the apis and filters test processes assert that OpenSSL reports FIPS-approved default properties and refuses MD5. The UBI 9 report stage runs the hash and signing tests the same way, so CI covers the provider-active path. Also: HmacError with a static message for the impossible tag length, stale test comments in sigv4.rs, the report description in docs/developing/fips.md, and test-fips-provider in make help. Signed-off-by: szedan <szedan@redhat.com>

szedan-rh · 2026-09-25T06:09:14Z

Wrapping up the second round, all in 22a8e47:

fips.md, relative OPENSSL_CONF path: doc now shows an absolute path and says a wrong path is silently a non-FIPS run.
fips.md, report description: a sha2/hmac finding that names aws-sigv4 now reads as "escaped dev-dependencies" instead of "add a line to FIPS_FEATURES".
signing.rs size: 513 lines implementation, 511 tests, most of the tests being the differential table against aws-sigv4.
sigv4.rs test comments: the S3 vector comment points at the aws-sigv4 crate that the differential test keeps this aligned with; assertion message says x-amz-content-sha256 is always emitted.
CI coverage: the UBI 9 report stage now runs the hash:: and aws:: tests on the FIPS provider, so make fips-check and the fips workflow exercise the provider-active path. Only those groups for now, the full FIPS-set run trips the 14 agentic_loop failures from feat(fips): enable certificate-only PostgreSQL Responses store #1335.
make help: added test-fips-provider (test-fips was already listed).
Provider state assertion: PRAXIS_TEST_FIPS_PROVIDER is set on cargo itself so it reaches the test binaries whatever the runner does; the apis and filters test processes then assert that OpenSSL reports FIPS default properties and refuses MD5. Checked the negative case, both fail without the provider.
Empty error stack on the impossible tag length (alexsnaps): replaced with a small HmacError carrying a static message.

Verified with clippy, make lint, make test-fips-provider and make fips-check on UBI 9 (31 tests on the provider, report clean).

Brings in AWS SigV4 signing through the system OpenSSL (praxis-proxy#1351) on top of the OpenAI Responses filter rename. Resolved docs/features.md, docs/fips.md, and Makefile by keeping both the renamed filter names and the AWS/FIPS additions. Signed-off-by: Sébastien Han <seb@redhat.com>

szedan-rh requested review from a team, alexsnaps and aslakknutsen September 24, 2026 14:56

leseb requested changes Sep 24, 2026

View reviewed changes

aslakknutsen reviewed Sep 24, 2026

View reviewed changes

fix(aws): keep Amazon's notice in signing.rs, fix Containerfile.fips …

09a2745

…comment Signed-off-by: szedan <szedan@redhat.com>

shaneutt assigned leseb Sep 24, 2026

shaneutt added the blocker This is blocking other work significantly label Sep 24, 2026

shaneutt added this to AI Gateway - Provider APIs + Agentic Loop Sep 24, 2026

github-project-automation Bot moved this to Next in AI Gateway - Provider APIs + Agentic Loop Sep 24, 2026

shaneutt added this to the v0.4.0 milestone Sep 24, 2026

shaneutt linked an issue Sep 24, 2026 that may be closed by this pull request

aws_sigv4_sign: FIPS-validated SigV4 signing (investigation) #814

Closed

3 tasks

szedan-rh requested review from aslakknutsen and leseb September 24, 2026 16:10

alexsnaps approved these changes Sep 24, 2026

View reviewed changes

shaneutt requested changes Sep 24, 2026

View reviewed changes

Comment thread docs/developing/fips.md Outdated

Comment thread docs/developing/fips.md Outdated

Comment thread filters/src/aws/signing.rs

Comment thread filters/src/aws/sigv4.rs

Comment thread Makefile Outdated

Comment thread Makefile

Comment thread Makefile Outdated

leseb approved these changes Sep 24, 2026

View reviewed changes

shaneutt self-requested a review September 25, 2026 00:16

Merge branch 'main' into fix_sigv4_fips

1a2650d

shaneutt approved these changes Sep 25, 2026

View reviewed changes

shaneutt merged commit 0048946 into praxis-proxy:main Sep 25, 2026
46 checks passed

github-project-automation Bot moved this from Review to Done in AI Gateway - Provider APIs + Agentic Loop Sep 25, 2026

Navigation Menu

Repository navigation

feat(aws): sign SigV4 requests through the system OpenSSL - #1351

feat(aws): sign SigV4 requests through the system OpenSSL#1351
shaneutt merged 4 commits into
praxis-proxy:mainfrom
szedan-rh:fix_sigv4_fips

szedan-rh commented Sep 24, 2026

Uh oh!

leseb left a comment

Uh oh!

aslakknutsen Sep 24, 2026

Uh oh!

szedan-rh Sep 24, 2026

Uh oh!

szedan-rh commented Sep 24, 2026

Uh oh!

alexsnaps left a comment

Uh oh!

alexsnaps Sep 24, 2026

Uh oh!

alexsnaps Sep 24, 2026

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

szedan-rh commented Sep 25, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

Conversation

szedan-rh commented Sep 24, 2026

Summary

Related issue

Validation

Breaking changes

Uh oh!

leseb left a comment

Choose a reason for hiding this comment

Uh oh!

aslakknutsen Sep 24, 2026

Choose a reason for hiding this comment

Uh oh!

szedan-rh Sep 24, 2026

Choose a reason for hiding this comment

Uh oh!

szedan-rh commented Sep 24, 2026

Uh oh!

alexsnaps left a comment

Choose a reason for hiding this comment

Uh oh!

alexsnaps Sep 24, 2026

Choose a reason for hiding this comment

Uh oh!

alexsnaps Sep 24, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

szedan-rh commented Sep 25, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants