| Titel | ggml-org llama.cpp (llama-server) 0.7.0-dev, build 50 Denial of Service |
|---|
| Beschreibung | A reachable assertion in the llama-server component of llama.cpp allows an attacker with access to the completion API to terminate the entire server process by submitting a crafted HTTP request. Exploitation does not require authentication when the endpoint is accessible without an API key or other access controls.
The server accepts the client-supplied grammar_triggers[].type integer and casts it to common_grammar_trigger_type without validating that it represents a supported enum value. An unsupported value, such as 99, subsequently reaches the default branch of a switch statement in common_sampler_init in common/sampling.cpp, triggering GGML_ASSERT and aborting the process with SIGABRT. The surrounding C++ exception handler does not intercept this process termination.
A single crafted POST request to /completion or /v1/completions can trigger the failure without requiring a grammar definition or special grammar configuration. The resulting process termination interrupts all active inference requests and makes the server unavailable until restarted. The reported path does not affect /v1/chat/completions, which removes grammar_triggers before processing.
The issue was reproduced on Linux x86_64 and independently reproduced on macOS arm64. The demonstrated impact is loss of service availability. |
|---|
| Quelle | ⚠️ https://github.com/ggml-org/llama.cpp/issues/30835 |
|---|
| Benutzer | Zyz3366 (UID 97230) |
|---|
| Einreichung | 22.10.2026 21:12 (vor 29 Tagen) |
|---|
| Moderieren | 21.11.2026 04:25 (29 days later) |
|---|
| Status | Akzeptiert |
|---|
| VulDB Eintrag | 441681 [ggml-org llama.cpp 0.7.0-dev llama-server common/sampling.cpp common_sampler_init grammar_triggers[].type erweiterte Rechte] |
|---|
| Punkte | 20 |
|---|