Knostic
-
Researchers find widespread remote code execution risk in AI inference engines from unsafe ZMQ and pickle use
Researchers found a recurring insecure pattern — pickle deserialization over unauthenticated ZeroMQ sockets — in multiple AI inference frameworks, creating remote code execution risks across projects including vLLM, NVIDIA TensorRT-LLM, Modular Max Server and SGLang; related research also showed browser and IDE injection risks in Cursor.
