An incomplete fix for a previous CVE left exception messages unsanitized in the vLLM inference engine. The Anthropic router handlers bypass the global FastAPI exception handler and directly construct JSON responses with str(exc), causing heap memory addresses from PIL.Image.open error messages to be exposed to clients. This leakage of internal memory addresses aligns with CWE‑532.

vLLM before version 0.23.1rc0 (vllm-project vllm library). The vulnerability impacts the Anthropic API router, the Server‑Sent Events streaming converter, and the realtime speech‑to‑text WebSocket components within vllm.

The CVSS score of 5.3 categorizes the flaw as moderate, while the EPSS score is not available and the vulnerability is not listed in CISA KEV. An unauthenticated attacker can trigger it by sending malformed image bytes to the Anthropic Messages API endpoints, which produces an UnidentifiedImageError that leaks a BytesIO repr containing a heap memory address. The exploitation requires no special privileges and results only in information disclosure, potentially aiding further attacks but not enabling code execution.