Supported Providers
VAAST is designed to scan your application's LLM API layer — the
endpoint your product code calls, not the underlying model provider's
infrastructure. If your app calls OpenAI internally and exposes a
/chat endpoint to users, point VAAST at your /chat endpoint.
That's where your system prompt, tool definitions, retrieval context,
and auth boundaries live — and where real vulnerabilities exist.
Direct scanning of model provider APIs (OpenAI, Anthropic, Google, xAI, Mistral, Cohere, etc.) is not supported. These endpoints are blocklisted. See Blocklisted Endpoints below.
How to configure your workspace
VAAST sends HTTP POST requests to your target URL with an OpenAI-compatible chat completions body:
{
"model": "...",
"messages": [{"role": "user", "content": "..."}],
"max_tokens": 512
}
Any endpoint that accepts this shape works with VAAST out of the box. Most LLM application frameworks (LangChain, LlamaIndex, Vercel AI SDK, FastAPI + LiteLLM, etc.) expose endpoints in this format.
Before scanning, you must verify domain ownership. See Domain Verification for details on DNS TXT, well-known file, API key, and written authorization methods.
Common target patterns
Your application endpoint
This is the primary use case. Point VAAST at the endpoint your users call:
- URL:
https://api.yourapp.com/v1/chat - Auth: Whatever your app requires (Bearer token, API key, etc.)
- Verification: DNS TXT or Well-Known File
Ollama (local)
- URL:
http://localhost:11434/v1/chat/completions - Body: OpenAI chat-completions. VAAST defaults to
"model": "llama3"— override by pointing the workspace URL at a target whose loaded model matches. - Auth: Usually none (local loopback).
- Verification: Localhost endpoints are auto-verified in VAAST — no domain verification step required.
Self-hosted HuggingFace TGI
If you are running your own HuggingFace TGI deployment
(self-hosted), point VAAST at your deployment's
/v1/chat/completions endpoint. This is not the HuggingFace
Inference API — that endpoint is blocklisted.
- URL:
https://your-tgi-host.example.com/v1/chat/completions - Body: OpenAI chat-completions (same format as Ollama).
- Auth:
Authorization: Bearer hf_...if the endpoint is gated. - Verification: DNS TXT or Well-Known File on your deployment domain.
Azure OpenAI
Azure OpenAI endpoints are your own deployments — fully supported.
- URL:
https://your-resource.openai.azure.com/openai/deployments/your-deployment/chat/completions?api-version=2024-02-01 - Auth:
api-key: your-azure-openai-keyin the workspace auth header. - Verification: Use the API Key verification method in VAAST Targets.
AWS Bedrock (via proxy)
Bedrock's native auth (SigV4) is not supported directly. Point VAAST at your application's endpoint that internally calls Bedrock. This is also the recommended approach — it tests the layer you own.
- URL:
https://api.yourapp.com/v1/chat(your proxy endpoint) - Auth: Your app's auth mechanism.
- Verification: DNS TXT or Well-Known File on your domain.
Blocklisted endpoints
The following provider API domains cannot be registered as VAAST scan targets. Attempting to add them will return a 400 error.
| Domain | Provider |
|---|---|
api.openai.com | OpenAI |
api.anthropic.com | Anthropic |
generativelanguage.googleapis.com | Google Gemini |
api.x.ai, x.ai | xAI / Grok |
api.mistral.ai | Mistral |
api.cohere.com | Cohere |
api.together.xyz | Together AI |
api.groq.com | Groq |
openrouter.ai | OpenRouter |
huggingface.co, api-inference.huggingface.co | HuggingFace Inference API |
These are the underlying model providers. VAAST is designed to test the application layer built on top of them — not the providers themselves.
Custom providers
For endpoints with non-standard body shapes, use the Request
Editor tab — it sends your raw body verbatim so you can hit any
HTTP API regardless of shape. A workspace-level body template with
{prompt} substitution for batch scans is planned but not yet wired.