Database/AI/ML frameworks & serving
vLLM (multimodal embeddings): Memory corruption
Impact
Memory corruption → crash and possible RCE
Who can reach it
Unauthenticated request to an exposed OpenAI-compatible serving port carrying crafted embeddings
What to do
Upgrade to 0.11.1+. If the provider hosts the endpoint, this is provider-owned; if the tenant runs their own vLLM, it is tenant-owned and the provider can only enforce network policy
Fleet impact
How widespread
Very common - vLLM is the default open-source inference server on nearly every neocloud's serverless-inference and BYO-endpoint product
Cost to remediate
daemon-restart - upgrade to 0.11.1+ and roll the serving fleet; every replica of every model endpoint must be restarted
Why it hits the whole fleet
The Completions API torch.loads user-supplied prompt embeddings; with PyTorch 2.8's sparse-tensor checks off by default a crafted tensor causes an out-of-bounds write - network-only, unauthenticated-adjacent, on every serving replica
References
Related entries
- OpenShift AI dashboard: incorrect network binding lets an in-cluster actor impersonate any user with an arbitrary tokenCVE-2026-16745 · Red Hat OpenShift AI odh-dashboard (network binding / token handling)High
- OpenShift AI training-operator: namespace edit/admin users escalate to host filesystem access via crafted training jobsCVE-2026-18982 · Red Hat OpenShift AI training-operator (ClusterRole aggregation + PodTemplateSpec passthrough)High
- vLLM (HF `auto_map`): Loads Hugging Face dynamic modules during model resolutionCVE-2026-22807 · vLLM (HF `auto_map`)High
- PyTorch (`weights_only` unpickler): Bypass of the `weights_only` allowlistCVE-2026-24747 · PyTorch (`weights_only` unpickler)High
- vLLM (hardcoded `trust_remote_code`, second instance): Same class, two more model files, through 0.18.0CVE-2026-27893 · vLLM (hardcoded `trust_remote_code`, second instance)High
- JupyterHub OAuthenticator: Authenticated user bypasses the intended identity checkCVE-2026-33175 · JupyterHub OAuthenticatorHigh
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.