A serious vulnerability has been discovered in AMD processors based on the Zen 5 architecture, posing a potential threat to the cryptographic integrity of affected systems. The flaw compromises the hardware implementation of the random number generator, making it possible to predict encryption keys. AMD has already begun distributing updates to address the issue, though full remediation is expected to continue into early 2026.
The vulnerability, identified as CVE-2025-62626 and catalogued by AMD under AMD-SB-7055, is linked to the RDSEED instruction, which is responsible for generating random numbers. In certain scenarios, an attacker with local access can force the instruction to return a zero value instead of a random output. This undermines the fundamental principles of cryptographic protection — data confidentiality and integrity. Moreover, the issue prevents systems from properly detecting generation failures: the process falsely registers as successful, potentially resulting in encryption keys composed entirely of zeros.
Only the 16-bit and 32-bit variants of the RDSEED instruction are affected; the 64-bit version remains fully functional and secure. The flaw was discovered in October by an engineer at Meta, prompting the Linux community to swiftly disable RDSEED across all Zen 5-based systems, issuing a patch through the Linux Kernel Mailing List. Notably, the disclosure was not made to AMD via the standard Coordinated Vulnerability Disclosure (CVD) process.
AMD has already released updated microcode for its Epyc 9005 (“Turin”) server processors, with patches for consumer-grade chips — including the Ryzen 9000, Ryzen AI Max 300, Threadripper 9000, and Ryzen Z2 series — expected by the end of November. Embedded Zen 5 variants are slated to receive updates in January 2026 through OEM partners.
Until official patches are deployed, AMD recommends several temporary mitigations: use only the 64-bit RDSEED instruction; disable RDSEED via the boot parameter clearcpuid=rdseed or an equivalent setting in QEMU; and discard any zero values returned, reissuing the command until a nonzero output is generated.
This is not the first time AMD has encountered issues with RDSEED. In 2021, a similar defect was identified in Cyan Skillfish APUs based on the Zen 2 architecture, where RDSEED consistently returned 0xffffffff. The solution at that time also involved disabling the instruction at the Linux kernel level while preserving functionality for RDRAND.
