CVE-2026-45667: open-webui: unauth endpoint drains embedding budget/DoS

### Summary GET `/api/v1/memories/ef` is accessible without authentication and executes `request.app.state.EMBEDDING_FUNCTION(...)`. This allows any unauthenticated caller to trigger embedding generation which can lead to direct cost exposure if a paid provider is used. Code reference: `backend/open_webui/routers/memories.py` (@router.get("/ef") -> calls `request.app.state.EMBEDDING_FUNCTION("hello world"))`. ### Details GET `/api/v1/memories/ef` is reachable without authentication and triggers request.app.state.EMBEDDING_FUNCTION("hello world"). This crosses an intended security boundary by allowing unauthenticated users to invoke potentially expensive embedding computation and/or paid upstream embedding APIs. ### PoC 1. Start Open WebUI in default configuration (no special env hardening; default ENABLE_MEMORIES is true). 2. From an unauthenticated client (no cookies/Authorization header), call: curl -i http://\<host\>:\<port\>/api/v1/memories/ef 3. Observe the server performs embedding generation and returns a response like: - HTTP 200 with JSON containing the result. How it can be abused / attacker actions: - Send repeated requests to `/api/v1/memories/ef` to: - consume CPU/GPU resources (DoS) - generate sustained outbound usage to embedding providers if configured (cost + rate-limit exhaustion) - degrade latency/availability for legitimate users ### Impact If embeddings are configured to use paid/remote providers (OpenAI/Azure/etc), an attacker can generate unlimited requests and incur charges. ## Resolution Fixed in commit [e5035ea31](https://github.com/open-webui/open-webui/commit/e5035ea31e179977e805a7032c979ff59a71860a), first released in **v0.8.0** (Feb 2026). The `/api/v1/memories/ef` route was removed entirely. It was a diagnostic/debug-style endpoint that hard-coded `"hello world"` through the embedding function without any authentication dependency; there was no legitimate caller that depended on it, so deletion was the cleaner fix than retrofitting auth. Users on `>= 0.8.0` are not affected.

An attacker discovers an Open WebUI instance running v0.7.2 via Shodan, automated port scanning, or by finding the instance in public security advisories. Without any credentials or prior access, the attacker writes a one-line loop: `while true; do curl -s http://target/api/v1/memories/ef; done`. Each request causes the Open WebUI backend to call `request.app.state.EMBEDDING_FUNCTION('hello world')` — invoking, for example, `openai.embeddings.create(model='text-embedding-3-small', input='hello world')` — generating a paid API call billed to the victim organization's OpenAI account at ~$0.00002 per call. At a modest 100 requests/second the attacker accumulates $0.002/sec; over 24 hours that reaches $172 — but with parallel workers or higher throughput the damage scales linearly. The attacker incurs zero cost. The victim detects the attack only when a cost alert fires or an engineer notices anomalous embedding usage in the provider dashboard.