Aurora PostgreSQL Babelfish flaw may let attackers run commands without a login

A heap overflow in Babelfish for Amazon Aurora PostgreSQL might let a remote attacker with no login run operating system commands as the database process user, the CVE record says. It takes a malformed packet sent to the TDS port, where Babelfish accepts connections from applications built for SQL Server. AWS published fixes on Nov 6 for Aurora clusters and the open-source extension, and CISA’s assessment in the record lists no exploitation.

Babelfish lets Aurora PostgreSQL understand TDS, the wire protocol SQL Server uses, so applications built for SQL Server can connect without changing drivers or code, AWS explains. AWS describes CVE-2026-119923 as a memory-safety flaw in how Babelfish handles the TDS connection handshake. The CVE record places it in the parser for the prelogin packet, as a heap-based buffer overflow, and folds in a related NULL-pointer flaw in the same parser.

AWS rates the flaw 9.8 under CVSS 3.1, with no privileges or user interaction needed, and its bulletin is marked “Important (requires attention)”. CISA’s assessment in the CVE record calls it automatable, with total technical impact. Neither AWS nor CISA reports attacks, and none of the sources mentions public exploit code. AWS doesn’t say who found the flaws, and the record lists no credits.

The same bulletin fixes CVE-2026-119924, an out-of-bounds read in the parser for TDS remote procedure calls. It needs a login, and it might let an attacker read adjacent process memory or cause a denial of service, the record says. AWS rates it 8.1, and CISA’s assessment calls it not automatable, with partial technical impact.

Only clusters with Babelfish turned on are affected, and an attack needs network access to the TDS port, 1433 by default. AWS tells customers to upgrade their clusters to the fixed version for their release and apply any pending maintenance. The self-managed Babelfish for PostgreSQL extension is affected in the versions AWS lists from 1.0 through 5.4, and AWS fixed it in 4.10, 5.6 and 6.1.