vLLM, an inference engine for large language models, had a flaw that caused exception messages to expose internal system data. The vulnerability stems from the conversion of FastAPI RequestValidationError objects to strings without proper sanitization, allowing unauthenticated users to trigger detailed error responses. The information included the operating system username, home and virtual environment paths, Python version, internal package structure, line numbers, and endpoint handler names. This weakness is classified under CWE-209, which deals with information exposure through error handling.

All installations of vllm prior to version 0.26.0 are affected. The flaw is present in the vllm/entrypoints/openai/server_utils.py and vllm/entrypoints/utils.py modules and affects the /v1/chat/completions, /v1/completions, /tokenize, and /detokenize endpoints. Upgrading to v0.26.0 or later removes the vulnerability.

The CVSS score of 5.3 indicates moderate severity. While the EPSS score is not available, the vulnerability can be exploited from any device that can send malformed JSON to the exposed endpoints, requiring no authentication. The lack of a KEV listing suggests no known widespread exploitation yet, but the internal data disclosed could assist attackers in planning further attacks. Organizations should treat this as a moderate risk that warrants prompt attention.