llama.cpp RPC server can be crashed or hijacked by unauthenticated peers
Anyone who can reach the llama.cpp RPC server over the network can send oversized data that overwrites memory on the server. This can crash it, and in some cases could let an attacker run their own code.
| Severity | CriticalCVSS 9.2 of 10 |
|---|---|
| Being exploited? | UnclearNo reports either way yet |
| Fix available? | Yesb11865 or later · since 28 Oct 2026 |
| How sure is this? | Official sourceBacked by an official source |
How the attack works simplified
- 1Attacker connectsReaches the RPC server over the network
- 2Sends HELLO messageClaims RDMA capabilities
- 3Sends oversized dataUp to 256 KiB, no authentication needed
- 4Stack overflowsWrites past a 1-byte stack variable
- ResultServer crashes or attacker may run code
What happened
An out-of-bounds stack write in the ggml-rpc RDMA transport function rdma_recv() lets unauthenticated peers send up to 256 KiB into a 1-byte stack variable in rpc_serve_client() after advertising RDMA capability in the HELLO message.
Are you affected?
You are affected if you run llama.cpp builds from b8841 up to (but not including) b11865 with the ggml-rpc RDMA transport and an exposed rpc-server.
llama.cpp b8841 before b11865 (ggml-rpc RDMA transport, rpc-server)
What to do
- Update llama.cpp to b11865 or later
- If you cannot update now, disable the RDMA transport or the rpc-server
- Do not expose the rpc-server to untrusted networks, and restrict access with a firewall
Timeline
- 28 Oct 2026, 03:17 UTCPublicly disclosed
- 28 Oct 2026, 03:17 UTCReported by NVD · NVD
- 28 Oct 2026, 03:22 UTCPicked up by Zero Radar
Sources 1 report, oldest first
Official vendor, government CERT or advisory database
- CVE-2026-118375NVD · 28 Oct 2026, 03:17 UTC
Technical details
- ID
- CVE-2026-118375
- CVSS
- 9.2 / 10
- EPSS
- 0.4% chance of exploitation within 30 days
- Disclosed
- 28 Oct 2026, 03:17 UTC
- On Zero Radar
- 28 Oct 2026, 03:22 UTC (5 min after disclosure)
- Last change
- 28 Oct 2026, 03:22 UTC
The plain-language parts and the attack diagram were written by an AI model (anthropic/claude-haiku-5.5) from the sources above. Facts like versions, exploitation and dates come from the sources themselves. How Zero Radar works