The ggml‑org/llama.cpp project released a fix on Oct 25, 2026. It adds a check that stops processing JSON deeper than 128 levels and returns a 400 error. Previously, extremely nested JSON could overflow the HTTP thread stack and crash the llama‑server (SIGBUS on macOS, SIGSEGV on Linux). The change applies to chat inputs, tool calls, and the /v1/systemone endpoint. The fix was co‑authored by Xuan Son Nguyen of Hugging Face.
Why it matters
Developers using llama.cpp can now rely on more stable server behavior when handling complex JSON payloads.
