Scanning Models¶

The LLMAdapter lets you scan any OpenAI-compatible model endpoint with zero agent code. It wraps the raw LLM in a minimal ReAct-style agent loop internally, providing it with tools and then testing whether it can be manipulated.

Basic Usage¶

from agent_redteam.adapters import LLMAdapter

adapter = LLMAdapter(
    base_url="http://localhost:8000/v1",
    api_key="your-key",
    model="your-model",
)

That's it. Pass this adapter to Scanner and run.

Configuration Options¶

adapter = LLMAdapter(
    base_url="http://localhost:8000/v1",
    api_key="your-key",
    model="your-model",
    # Optional parameters:
    system_prompt="You are a helpful coding assistant.",  # Custom system prompt
    max_turns=10,        # Max ReAct loop iterations (default: 10)
    temperature=0.1,     # LLM temperature (default: 0.1)
    timeout=60.0,        # Per-request timeout in seconds (default: 60)
)

Parameters¶

Parameter	Type	Default	Description
`base_url`	`str`	required	OpenAI-compatible API base URL
`api_key`	`str`	required	API key for authentication
`model`	`str`	required	Model name/identifier
`system_prompt`	`str`	built-in	System prompt for the agent wrapper
`max_turns`	`int`	`10`	Maximum tool-use turns per task
`temperature`	`float`	`0.1`	Sampling temperature
`timeout`	`float`	`60.0`	HTTP request timeout (seconds)

How It Works Internally¶

The LLMAdapter constructs a minimal agent loop around your model:

flowchart TD
    Task[Attack Task] --> SystemPrompt[System Prompt + Tool Descriptions]
    SystemPrompt --> LLM[Your Model]
    LLM --> Parse{Tool Call?}
    Parse -->|Yes| ExecuteTool[Execute Tool]
    ExecuteTool --> LLM
    Parse -->|No| FinalOutput[Final Response]
    FinalOutput --> Trace[AgentTrace]

The model receives a system prompt listing available tools
It can call tools by outputting {"tool": "name", "args": {...}} (bare or in markdown code blocks)
Tool results are fed back as the next user message for the next turn
Every turn is captured into the AgentTrace: system prompt, user messages, each model response (with token counts), tool calls with arguments, tool results, and the final output — giving full visibility for detectors and post-scan analysis

Compatible Providers¶

Any provider exposing the OpenAI chat completions API works:

Provider	Example `base_url`
OpenAI	`https://api.openai.com/v1`
vLLM	`http://localhost:8000/v1`
Ollama	`http://localhost:11434/v1`
Azure OpenAI	`https://{name}.openai.azure.com/openai/deployments/{model}`
Together AI	`https://api.together.xyz/v1`
Any OpenAI-compatible	`http://your-endpoint/v1`

When to Use LLMAdapter vs CallableAdapter¶

Use Case	Adapter
Testing a raw model's safety guardrails	`LLMAdapter`
Testing how a model handles tools	`LLMAdapter`
Testing your custom agent with specific logic	`CallableAdapter`
Testing an agent with custom tool implementations	`CallableAdapter`

LLMAdapter is the lowest-barrier entry point — if you can curl a model endpoint, you can scan it. For testing agents with custom business logic, routing, or memory, use CallableAdapter instead.

Example: Scanning a vLLM Model¶

See examples/vllm_agent_scan.py for a complete working example that scans a vLLM-hosted model across 8 vulnerability classes with detailed trace output showing the full model conversation, tool calls, and signal detection.