GPU VulnDB

Database/Kernel, userspace & hypervisor

Linux kernel crypto: AEAD algorithms using cra_reqsize get an uninitialized reqsize, corrupting memory

CVSS 9.8CVE-2025-68726Kernel, userspace & hypervisorcurated

Impact

The cra_reqsize field added to crypto_alg was wired up for ahash and acomp only, but drivers were being told to use it generally. An AEAD algorithm that sets cra_reqsize instead of calling crypto_aead_set_reqsize ends up with a request size the framework never propagates, so requests are allocated too small and the algorithm writes past them - memory corruption and crashes in kernel context. Whether a node is exposed depends entirely on whether a loaded AEAD implementation uses the new field, which in practice means specific hardware crypto drivers rather than the generic software ones. On a GPU node the AEAD path is reached through dm-crypt, IPsec and kernel TLS, so the practical consequence is a crash on a node that is expensive to drain rather than a tenant-reachable privilege escalation. The fix adds the proper set_reqsize calls in the AEAD init path.

Who can reach it

Local, through any kernel consumer of an affected AEAD transform (dm-crypt volume, IPsec SA, kTLS socket). No attacker interaction is needed for the corruption if an affected algorithm is in use - normal traffic exercises it. Reachability requires an AEAD driver that sets cra_reqsize to be loaded.

What to do

Take the stable-kernel fix (commits below) and reboot the node; this is core crypto API code with no runtime workaround. Drain GPU workloads first. Before scheduling, check whether the kernel you run actually loads an AEAD driver that sets cra_reqsize - if not, the node is not reachable by this and the reboot can wait for the next window.

References

Related entries

All Kernel, userspace & hypervisor entries

This entry is curated: imported from vendor advisories with machine assistance, not yet individually verified. Confirm against your vendor's advisory before acting, and report anything wrong.