CVE-2026-78147 exposes a dangerous pattern in llama.cpp's GGML-RPC deserialization: the earlier CVE-2026-34159 patched a null-pointer trigger in create_node() without addressing the underlying op/op_params validation gap. These parameters control what operation gets instantiated during deserialization, and when passed through without schema enforcement, they create a direct injection surface into the inference pipeline.
The 7.3 CVSS score undersells the risk. GGML-RPC is a distributed protocol—servers processing remote data with inadequate input validation are exactly the conditions for remote exploitation. An attacker who can reach the RPC endpoint can potentially manipulate what operations a remote inference node executes, and because RPC coordinates multiple machines, a successful injection may poison shared state across the entire inference cluster.
Audit your deployment immediately: verify that ggml-rpc.cpp applies explicit validation to op and op_params before they reach any instantiation logic—do not rely on the CVE-2026-34159 null-pointer guard alone, as that was a symptom patch, not a structural fix. Check whether downstream code in your inference pipeline has implicitly trusted unvalidated RPC commands, because adding proper validation now may break assumptions elsewhere.
The automatic closure of the original GitHub issue is a red flag—vulnerabilities punted due to inactivity frequently resurface as follow-on CVEs. Review other create_node() callers and deserialization entry points in GGML-RPC for similar unvalidated op/op_params patterns; where one path was partially patched, structurally similar paths often carry the same weakness未被触及.