| Title | SciPhi-AI R2R <=3.6.6 Server-Side Request Forgery |
|---|
| Description | R2R through 3.6.6 is vulnerable to server-side request forgery in POST /v3/retrieval/completion and POST /v3/retrieval/agent. The client may set generation_config.api_base. Models routed to LiteLLM (prefixes other than openai/, azure/, deepseek/, ollama/, lmstudio/, anthropic/, azure-foundry/, or an unprefixed name) cause the server to send HTTP requests to that URL. Default Docker configuration sets require_authentication to false. Whether OPENAI_API_KEY is attached depends on the LiteLLM provider: huggingface/ performs the SSRF without that header; OpenAI-compatible prefixes such as custom/ (and together_ai/ against a custom api_base in tested LiteLLM builds) send Authorization: Bearer <OPENAI_API_KEY> to the attacker-controlled host. Models with the openai/ prefix are handled by OpenAICompletionProvider, which ignores api_base and are not this issue. |
|---|
| Source | ⚠️ https://gist.github.com/DReazer/6c1fd0ef98900fbbfa383faa20f731f6 |
|---|
| User | Snkn0w (UID 90071) |
|---|
| Submission | 08/26/2026 07:57 (1 month ago) |
|---|
| Moderation | 10/03/2026 20:05 (1 month later) |
|---|
| Status | Accepted |
|---|
| VulDB entry | 413373 [SciPhi-AI R2R up to 3.6.6 Retrieval Completion API Endpoint llm.py generation_config.api_base server-side request forgery] |
|---|
| Points | 20 |
|---|