GPU VulnDB

Database/AI/ML frameworks & serving

SGLang: unauthenticated PUT /route poisons the KV transfer routing table in disaggregated mode

CVSS 8.8CVE-2026-92972AI/ML frameworks & servingcurated

Impact

In prefill/decode disaggregation mode the prefill bootstrap service exposes an unauthenticated PUT /route endpoint that accepts arbitrary rank_ip and rank_port values. Anyone who can reach that port can point decode workers at an endpoint they control, which stops serving for the affected ranks and leaks KV transfer metadata - session identifiers and tensor-parallel topology parameters. Disaggregated deployments are exactly the large multi-node setups where a single poisoned route can idle a whole block of GPUs, and the bootstrap port is usually treated as internal plumbing rather than an authenticated API, so it is often bound wider than intended. Recovery means restarting the affected workers, which drops their KV cache and in-flight requests.

Who can reach it

Anyone with network reach to the prefill bootstrap service port - any tenant pod on the same cluster network, or anything on the inference VLAN if the port is not firewalled. No authentication required.

What to do

No fixed version is named in the record (affected through 0.5.19); track the upstream issue. Until a release lands, restrict the bootstrap port to the serving nodes themselves with network policy or host firewall rules and keep disaggregation traffic off any tenant-reachable network. Applying a fix or restricting the port is a serving-process restart, not a node reboot.

References

Related entries

All AI/ML frameworks & serving 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.