[!NOTE] This report was written by Claude (Anthropic's Claude Code) under the direction of @lukebakken, who reviewed it before filing. The analysis, the proposed fix and its tests were produced by Claude and verified against
mainbefore submission.
amqp10_client, the Erlang AMQP 1.0 client library published on Hex.pm and vendored in rabbitmq-server as deps/amqp10_client.
amqp10_client_connection:format_status/1 is the gen_statem status redactor: OTP calls it before the process state is written to a crash dump, and it also runs on sys:get_status. It delegates to obfuscate_state_config_sasl/1, whose only redacting clause matches {plain, User, Pwd}.
The connection state never holds that shape. open/1 runs the config through maybe_encrypt_sasl/1, which converts {plain, User, Pwd} into credentials_obfuscation:encrypt(term_to_binary(Plain)), so #state.config holds an encrypted_sasl() — {plaintext, Bin} or {encrypted, Bin}. The decrypted tuple exists only in transient locals during the handshake. The redacting clause is therefore unreachable and the Other -> Other fallback returns the credential-bearing term unchanged.
credentials_obfuscation:encrypt/1 returns {plaintext, Bin} whenever no obfuscation secret is configured, and no secret is configured by default. In that mode the stored value is {plaintext, term_to_binary({plain, User, Pwd})}, and the password is recoverable from it by inspection.
This is not that issue, and not a duplicate of it, but they are the same defect class and a reader will reasonably ask.
CVE-2026-67406 (GHSA-q44f-9vc5-grh7, patched 4.3.3 / 4.2.9 / 4.1.14 / 4.0.23) reported that the shovel gen_server processes implemented no format_status callback at all, so the crash reporter wrote their deobfuscated plaintext URIs to the error log. Its own text notes that grep -rn format_status across the shovel source returned zero results. The fix, a7cc6b5b50, added format_status/1 to rabbit_shovel_worker.
That commit did not touch amqp10_client_connection, and the module's history since is unrelated to redaction. An AMQP 1.0 shovel starts an amqp10_client_connection process holding the same credentials, in the same supervision tree, and that process's redactor has been ineffective since 2023 — predating CVE-2026-67406 by two years. So the shovel worker is now redacted while its client connection is not.
The two differ in substance as well as in scope: CVE-2026-67406 concerned a missing callback exposing plaintext URIs, this concerns a present callback matching a shape the state never holds. They also affect different packages — that one rabbitmq-server, this one the amqp10_client Hex.pm package, which has consumers well beyond RabbitMQ.
Standalone users of amqp10_client are the affected population. Nothing in the library sets an obfuscation secret, so the default is {plaintext, ...} and a crash dump or a captured sys:get_status carries the broker password in recoverable form.
RabbitMQ brokers are materially less exposed, and less so than CVE-2026-67406 was. rabbit_prelaunch_dist:set_credentials_obfuscation_secret/0 always calls credentials_obfuscation:set_secret/1 with the Erlang cookie, and enabled defaults to true, so on a node the unredacted term is PBE ciphertext rather than a password. That is weaker than intended — anyone able to read a crash dump can often read the cookie beside it — but unlike CVE-2026-67406, which exposed deobfuscated plaintext URIs, it is not direct credential disclosure. Within the broker the path is reachable through rabbitmq_shovel, which builds an AMQP 1.0 connection config from a shovel URI via amqp10_client:parse_uri/1 (rabbit_amqp10_shovel.erl), and through rabbitmq_amqp_client.
Reading the data requires access to the node's crash dump or an Erlang shell on it, so this is a defence-in-depth failure in a redactor rather than a remotely reachable flaw. The redactor exists precisely because that material is expected to be handled or shipped less carefully than live process memory.
deps/amqp10_client/src/amqp10_client_connection.erl:
obfuscate_state_config_sasl(Cfg) ->
Sasl0 = maps:get(sasl, Cfg, none),
Sasl = case Sasl0 of
{plain, Username, _Password} -> %% never matches
{plain, Username, <<"[redacted]">>};
Other ->
Other %% credential passes through
end,
Cfg#{sasl => Sasl}.
The sibling obfuscate_state_config_tls_opts/1 in the same redactor correctly strips password from the TLS proplist, so this is per-branch drift rather than a deliberate scope decision.
format_status/1 and this redactor. At that commit open/1 did not encrypt, so the stored value really was {plain, User, Pwd} — the redactor was correct as written.credentials_obfuscation, adding encrypt_sasl/1 and the encrypted_sasl() type. It changed the stored shape and did not update the redactor.The earliest release tag containing 2316579c40 is v3.13.0.
Redact by allowlist rather than by enumeration, so that a future SASL representation is redacted rather than passed through:
Sasl = case Sasl0 of
none -> none;
anon -> anon;
external -> external;
{plain, Username, _Password} ->
{plain, Username, <<"[redacted]">>};
{plaintext, _} ->
{plaintext, <<"[redacted]">>};
{encrypted, _} ->
{encrypted, <<"[redacted]">>};
_ ->
{plaintext, <<"[redacted]">>}
end,
Enumerating the two encrypted_sasl() shapes alone would fix the observed leak but preserve the failure mode that caused it: the denylist shape is what allowed an unrelated change to reintroduce the leak silently.
A commit implementing this, with two eunit cases, is ready to push to a private fork on request. The first asserts that the password does not appear anywhere in term_to_binary/1 of the whole formatted status, rather than checking the sasl field, and asserts the password is recoverable before redaction so it cannot pass vacuously. Reverse-applying the fix makes it fail. gmake dialyze and gmake xref are unchanged against main.
Not runtime-confirmed against a live crash dump. Verified by source inspection on main and by unit test: with no obfuscation secret set, maybe_encrypt_sasl({plain, User, Pwd}) returns {plaintext, term_to_binary({plain, User, Pwd})}, and format_status/1 returns a state from which that term is recoverable verbatim.
Found while reviewing static-analysis findings against the 4.3.6 source.