AMD's Instinct bulletin logs a corrected CVE, and the number it dropped has no record

AMD's Instinct bulletin logs a corrected CVE, and the number it dropped has no record

AMD's security bulletin AMD-SB-6033 now gives a single CVE identifier for the flaw it describes in the ROCm Communication Collectives Library, and says so itself. The bulletin's revision entry dated 2026-10-06 reads "Corrected CVE in the product table", under an entry dated 2026-09-30 that reads "Initial version".

A close-up of chip dies on a silicon wafer, their circuitry diffracting light into red, green and yellow bands

Diffracted light on an unpackaged wafer, photographed for AMD's developer site. Image: AMD

Read at 22:53 UTC on October 6, the affected-products table has nine rows, one per AMD Instinct accelerator, and every one of them reads CVE-2026-43598. Counting the string across the whole page gives 10 occurrences, the nine product rows plus the single row of CVE details, and no occurrence at all of CVE-2026-54974.

Until that revision those nine rows read CVE-2026-54974, a number that has never had a published CVE record. A request to the CVE Program's record service for CVE-2026-54974 returned "The cve record for the cve id does not exist" at 22:53 UTC on October 6.

Two known-good identifiers were queried in the same run as a control. CVE-2024-3094, the xz backdoor record filed by Red Hat, and CVE-2026-0482, an AMD record for a Versal secure boot bypass, both came back as full published records in the same minute, so the service was answering normally.

The surviving identifier now has a record of its own

CVE-2026-43598 is in state PUBLISHED, with a datePublished of 2026-10-06T20:06:56.948Z and a dateReserved of 2026-05-01T18:15:55.867Z. AMD is the assigning authority on it, which is what you would expect of a company that files its own records as a CVE Numbering Authority.

The record and the bulletin agree on the hardware. Both name the same nine accelerators, the Instinct MI210, MI250, MI250X, MI300A, MI300X, MI308X, MI325X, MI350X and MI355X, each with ROCm 7.14 marked unaffected, and both carry the CVSS 4.0 vector CVSS:4.0/AV:N/AC:H/AT:N/PR:L/UI:N/VC:H/VI:H/VA:H/SC:N/SI:N/SA:N for a base score of 7.7.

They differ on two things. The record classifies the flaw as CWE-822, Untrusted Pointer Dereference, a weakness class AMD's bulletin never gives, while the bulletin thanks the researcher Luna Nova of lunnova.dev and the record carries no credit to anyone. Its only reference is the bulletin itself.

Both documents describe the flaw the same way, as "improper input validation" in RCCL that "could allow a compromised peer rank or network-adjacent attacker to dereference an attacker-controlled pointer, potentially resulting in remote code execution". The wording has not moved since September 30, and neither has the 7.7 score.

Nothing here reaches a desktop

This is datacentre silicon from end to end. RCCL is the ROCm library that shuffles data between accelerators inside a server, AMD places the fault in its proxy communication path, and the nine products named are Instinct parts that go into racks.

No Radeon card, no Ryzen processor and no Adrenalin driver is among them, and nothing in the bulletin asks a consumer to install anything. The mitigation AMD names is ROCm 7.14, with a mitigation release date of 2026-07-15, and ROCm runs on the accelerators above rather than on anything in a gaming PC.

AMD's index of security bulletins, read at 22:53 UTC on October 6, dates AMD-SB-6033 to September 30 with a last updated date of October 6, and lists CVE-2026-43598 against it and nothing else. Across all 207 bulletins in that index, CVE-2026-54974 appears nowhere.

AMD has said nothing beyond the one line in its revision log. It has not explained how the wrong identifier came to be published nine times over, and it has not revised the description or the score, so the flaw an operator is being asked to patch is the one it was on September 30.

Leave a Comment

Your email address will not be published. Required fields are marked *