pgEdge Go LLM Library health endpoint pings providers with the operator's real credentials

CVE-2026-125328 (CWE-200/CWE-522) is a credential-exposure defect in pgEdge's Go LLM Library. From 0.1.0 until 0.2.0, the proxy's GET /v1/health handler pings LLM providers using the operator's real credentials, so those credentials are exercised and could be reflected/abused through the health endpoint. Fixed in 0.2.0.

This is an AI LLM-proxy helper that holds the operator's provider API keys; the health handler that presents those credentials is an outbound channel that, if exposed, lets an attacker probe or exfiltrate the operator's provider identity and turn the health route into a server-side request primitive against the LLM provider.

The proxy's GET /v1/health handler (handleHealth) pings configured LLM providers to report health and reuses the operator's real provider credentials to do so. If the health endpoint is exposed/accessibly reachable, the operator's credentials are used or surfaced on every health check, and the endpoint can be abused as a server-side request channel.

pgEdge Go LLM Library (pgedge-go-llm-lib) from 0.1.0 until 0.2.0 — llm/proxy/handlers.go GET /v1/health.

Upgrade pgedge-go-llm-lib to 0.2.0 or later (commit 7531a55d368b3a29cd173c792cc4609c4134b076); do not expose the health endpoint to untrusted callers, and avoid echoing provider/credential details in health responses.