Release notes for IRRD 4.4.8

IRRD 4.4.8 was released on October 19, 2026, and fixes several security issues, including two in RPSL change submission processing:

  • Unbounded password verification work in RPSL submissions (GHSA-pcwp-hhmm-c58f). This affects IRRD 4.3.0 and all later versions.

  • Database connection leak on errors in RPSL submission handling (GHSA-cvm7-qcww-5m59). This affects IRRD 4.4.0 and all later versions.

Both issues were discovered and reported by T-Chachamaru. All users of 4.4 are recommended to upgrade.

Unbounded password verification work in RPSL submissions

IRRD did not limit the number of passwords in an RPSL change submission, verified each password against every auth hash of each relevant mntner, and did not limit the cost factor of BCRYPT-PW hashes in mntner objects. An attacker able to create or modify a mntner in an authoritative source could store a bcrypt hash with an extreme cost factor, after which anonymous submissions referencing that mntner could each consume very large amounts of server CPU time, degrading or denying the processing of RPSL submissions. The impact is limited to availability.

IRRD now rejects newly submitted BCRYPT-PW hashes with a cost factor over 16, and evaluates at most 20 distinct passwords or override passwords per submission. Note that:

  • Hashes stored before upgrading are not affected by the new limit. Operators can audit existing objects for BCRYPT-PW hashes with a cost factor over 16.

  • Existing mntner objects with such a hash must have it replaced before other modifications to the object are accepted.

Database connection leak on errors in RPSL submission handling

A crafted RPSL change submission, for example through the HTTP API, could trigger an unhandled error in IRRD, after which the PostgreSQL connection used by that request was not returned to the connection pool. Internal caches could retain these connections, gradually reducing the connections available to the affected worker and to other users of the same PostgreSQL instance, degrading IRRD’s database-backed functionality. The known trigger is an api-key attribute that is not a valid UUID. No authentication was required. The impact is limited to availability, and the effect recovers after a restart.

IRRD now releases database connections on all error paths, in submission processing and several other places, ignores API keys that are not valid UUIDs, and no longer retains request state in caches.

Other changes

  • The number of fields a GraphQL request may ask for is now limited, over all queries in the request, including through aliases and fragments, which could multiply the work performed. This was a hardening measure also suggested by T-Chachamaru.

  • The expected network level protections for public IRRD instances are now covered in the deployment documentation, and the handling of availability issues in the security policy.

  • Users with user management permission on mntners in several sources could read change log entries for one of their mntner names in another of their sources, where they do not have permission on that mntner. This also required knowing the randomly generated UUID of the entry, which IRRD only discloses in the change log of a mntner on which the user does have permission, so this is rather theoretical. This affects IRRD 4.4.0 and all later versions.

  • Reuse of a previous one time password now results in a form validation error, rather than an internal server error. This affects IRRD 4.4.0 and all later versions.

  • The recipient of all outgoing e-mail must now be a single address, and its subject is unfolded before sending. As the From header of a submission is a mailbox-list, and the envelope recipients of a reply are taken from its To header, a submitter could previously have the reply delivered to additional addresses. Submissions whose From is not exactly one address are still processed, but receive no reply. This affects all IRRD 4 versions.

  • Migrating a maintainer is refused when no e-mail address can be found for any admin-c contact in the same source, rather than starting a migration that can not be completed.