Hardware-Aware AutoML EUVD-2026-56531

Exploitation requires three concurrent conditions: (1) the attacker must have local access to the host running the AutoML software and must be able to write a file into at least one directory that appears in the application's library or executable search order ahead of the legitimate file location - this is the AT:P (attack requirements present) condition in the CVSS 4.0 vector; (2) a high-privileged user (administrator, root, or equivalent) must execute the Hardware-Aware-Automated-MachineLearning NA application after the malicious file has been planted - this is the UI:P (passive user interaction) and PR:H condition; and (3) the affected software version must be earlier than commit 45cd723. Exploitation is limited to local access (AV:L) - remote unauthenticated exploitation is not possible. There is no subsequent-system scope change (SC:N/SI:N/SA:N), meaning the impact is confined to the vulnerable host. The requirement for a high-privileged user to independently trigger execution, combined with the need for a writable directory in the search path, substantially narrows the viable attacker population compared to a straightforward network-exploitable flaw.

The CVSS 4.0 base score of 5.4 (Medium) accurately reflects the constrained threat model: local access, high privilege required (PR:H per the vendor-supplied vector), specific attack prerequisites (AT:P), and passive privileged-user interaction (UI:P) must all align simultaneously. No EPSS data was provided in the intelligence package, and the vulnerability is absent from the CISA KEV catalog, indicating no confirmed active exploitation. No public proof-of-concept code has been identified. A meaningful discrepancy exists between the CVSS PR:H metric and the description's characterization of an 'unprivileged software adversary' - the vendor vector appears to assign PR:H to reflect that a high-privileged user must be present to trigger the load, not that the attacker themselves requires high privileges; this interpretation should be verified against the Intel SA-01485 advisory. In practice, exploitation requires physical or logical access to the target host plus the ability to write into a directory on the application's search path, which meaningfully limits the attacker pool to already-local principals. For organizations running this specialized AutoML framework in shared or multi-tenant development environments, the risk is higher than the CVSS score alone suggests, because such environments frequently have permissive interim directory permissions.

An attacker with a local unprivileged shell on a shared ML development server identifies that the AutoML framework's launch wrapper resolves a shared library by scanning a world-writable temporary directory before the system library path. The attacker compiles and drops a malicious shared object into that directory under the expected library name. When a data scientist or administrator subsequently runs a hardware-aware NAS job - passive interaction requiring no explicit co-operation - the AutoML process, executing under the high-privileged user's token, loads the attacker's library and the attacker gains code execution at the elevated privilege level. No public proof-of-concept code has been identified, but the technique is well-understood and requires no specialized knowledge once the writable path position is confirmed.