CVE-2026-34175 is a DLL search order hijacking vulnerability in HA-AutoML tooling. The component-level impact reads high across confidentiality, integrity, and availability—but the subsequent system impact rating of 'none' is where your attention is needed, because it's likely misleading for this class of tool.
HA-AutoML systems don't operate in isolation. They execute on developer workstations or HPC nodes where they hold active connections to GPU drivers, CUDA runtimes, hardware accelerators, and model artifacts containing sensitive IP. When you compromise the DLL search path of such a tool, you're not just escaping that single process—you're inheriting the trust relationships it already established with critical system resources. This is what makes HA-AutoML a trust broker, not an island. The 'none' subsequent impact rating treats the blast radius as contained within the component's own scope, but that assumption fundamentally mismodels how trust-brokers behave once compromised.
Two additional factors make this more severe than the CVSS suggests. First, DLL hijacking is a persistent condition, not a one-shot exploit. Once a malicious DLL lands in a trusted search path location, every subsequent invocation of the HA-AutoML tool re-triggers code execution without additional attacker action. The blast radius doesn't close—it stays open until the DLL is found and removed, which in development environments could be months. Second, the 'passive user interaction' requirement is thinner than it appears. HA-AutoML pipelines routinely pull external dependencies, checkpoint models, and load community-contributed components—scenarios where an unprivileged adversary could pre-poison a search path location that the tool trusts.
What to check: inventory which HA-AutoML tools run on sensitive workstations or HPC nodes in your environment. Audit their DLL search paths for writable locations an unprivileged user could plant into. Treat any confirmed DLL hijack in this context as a trust broker compromise—remediate it with the urgency you'd apply to local privilege escalation, not as a low-severity path manipulation. The 'none' system impact rating is a scoring convention artifact, not a signal that lateral movement pathways don't exist.