Skip to content

Security review, November 2026

Item Value
Date 2026-11-18
Reviewer Sebastien Rousseau, with AI-assisted analysis (Claude)
Version reviewed main at 21782e8 (0.0.40 plus the 0.0.41 branch so far), and the fixes below
Against The assurance case: requirements R1 to R5, the trust boundaries and the weakness table
Result Three low-severity defects found and fixed (F-01, F-02, F-09); six findings accepted with reasons

Scope

  • Library code: every class in src/main/java/co/cryptoservice/sdk/, with most attention on CryptoServiceClient (builder and validation), Transport, Exchange, Backoff, Route, Operation, Decoder, Json, Problem, ApiError, TransportError, Redacted, Traceparent, RawResponse and ValidationError.
  • Transport and TLS: the HttpClient the SDK builds, the injected client path, redirects, the scheme, timeouts and the deadline.
  • Credential handling: where the API key and JWT are held, which requests carry them, and every string that could print them.
  • Input parsing: problem bodies, success bodies, Retry-After, Content-Type, and caller input to the builder and the raw request.
  • Error paths: what each failure raises, what its message says, and whether a future can be left uncompleted.
  • CI workflows and permissions: all nine workflows under .github/workflows/.
  • Release pipeline: release.yml, the tag checks, the release environment, signing and the published assets.
  • Dependency set: the runtime dependency tree and the build plugins.

Out of scope: crypto-server and its cryptography (reviewed in crypto-service), GitHub's and Maven Central's own infrastructure.

Method

  1. Read every class in scope line by line against the assurance case, following each entry point to where it leaves the SDK.
  2. For each weakness row of the assurance case, found the code and the test that support it, and tried to break the claim: for example a problem type holding CR/LF, a raw path of //evil.example/x, a Retry-After of 1.2.3, a closed HttpClient under a pending retry (the JDK returns a failed future, so the call still completes).
  3. Wrote a failing test for each suspected defect before changing code; kept a fix only once the test passed with it.
  4. Checked every workflow for its permissions, pinning, checkout credentials and expression use in run: steps; actionlint reports nothing.
  5. Ran make audit (OSV-Scanner over the CycloneDX SBOM, 16 packages including test scope): no known vulnerability. Fuzzed the new RetryAfterFuzzTest for 60 seconds with no finding; the other two targets' seeds replay in every build and CI fuzzes them on every pull request.
  6. Rebuilt the 0.0.40 release and compared it with what was published (reproducing a release).

Findings

Severity is for a caller of the SDK: High (secret disclosure or integrity loss reachable by a network attacker), Medium, Low (needs a hostile server or an unusual configuration, limited impact), Info (hardening, no defect).

ID Severity Finding Disposition
F-01 Low Log forging through problem members (CWE-117). ApiError's message and Problem.toString quoted the server's type, title, instance and code verbatim. A hostile server could put CR/LF or U+2028 in them and forge lines in any log that records the exception, which the SDK documents as safe to log. Fixed in 51da109: Redacted.printable escapes control characters and line separators in every shown value and in the message. ErrorModelTest.serverTextCannotForgeLogLines failed before the fix; ProblemFuzzTest now asserts no control character reaches either string.
F-02 Low The same through ValidationError.toString (CWE-117). The record kept its default toString, printing the server's field and message verbatim. Fixed in c6ff4af: it now uses Redacted.show. The new assertion failed on the default toString.
F-03 Low Response bodies are buffered whole (CWE-770). Every response is read with BodyHandlers.ofByteArray(), with no size limit, so a hostile or broken server can make a call hold an arbitrarily large body in memory. Jackson's stream read constraints bound nesting and token sizes, not the body. Accepted for now, on the roadmap: the server is the caller's own deployment, inside its trust for availability, and a limit needs an API decision (a default and a builder option).
F-04 Low A slow body is bounded only by the deadline. The per-attempt timeout covers connecting and the response headers, as its Javadoc says; a body that trickles in never times out unless the caller sets deadline or cancels the future. Accepted, documented: deadline and cancellation exist for this, and the Javadoc of timeout says what it covers.
F-05 Low Cleartext http is allowed to any host (CWE-319). The builder accepts http:// for any host, so a caller can send credentials in cleartext off loopback. Accepted, documented (README Security, SECURITY.md): loopback and in-cluster sidecar deployments use http, and refusing it would break them. Revisit with the suite (a loopback-only default would be a breaking change).
F-06 Info The release build job runs the full gate with id-token: write and attestations: write. Any build plugin or test runs while the job can mint an OIDC token. Accepted: every plugin and dependency is pinned by SHA-256 and every action by commit; the Central token and GPG key are not in that job. Moving the attestation to a job that only hashes the uploaded assets would narrow it; listed as a possible task.
F-07 Info The tag signature is checked against the KEYS.asc of the tagged tree. Someone able to push to main could add a key and sign a tag with it in the same change. Accepted: main takes only signed commits through a pull request with the required checks, and publishing also needs the release environment's required reviewer.
F-08 Info Credentials are held as Strings in Transport for the client's lifetime and cannot be wiped. Accepted: the JDK's HttpRequest headers are strings anyway; the SDK never prints them (ClientTest.builderToStringRedactsCredentials).
F-09 Low make api-check passed when it could not fetch its baseline. It discarded curl's error and treated an empty answer as "no release yet", so a network or Central failure turned the API breakage gate into a pass. Fixed in 1f6af75: it now fails; make api-check-selftest checks both that a breaking change and a missing baseline fail it, and it failed on the old recipe.

No finding reached Medium or High. Checked and found sound, with the tests that hold them: redirects refused (ClientTest.redirectsAreRefused), no TLS bypass, base URL and header validation, credentials only on authenticated routes, credentials never in a message or toString (ModelTest), the retry-safety table, bounded retries and Retry-After, deserialisation into fixed records only, a raw path cannot change the host (ClientTest.rawPathCannotChangeTheHost, added by this review), and no expression interpolation in workflow run: steps.

Follow-up

  • F-03: a response size limit, on the roadmap.
  • F-06: a separate attestation job, recorded as a possible task.
  • The next review is due with the first release after 2027-05-18, or sooner if a trust boundary changes.