The attacker and judge LLMs run on a provider you choose. Set the provider’s API key in your shell or a .env file before running, then reference it by name in attackerLlm.apiKeyEnv.
Supported providers
| Provider | provider value | Env var | Default model | Notes |
|---|
| OpenAI | openai | OPENAI_API_KEY | gpt-4o-mini | |
| Anthropic | anthropic | ANTHROPIC_API_KEY | claude-3-5-haiku-20241022 | |
| Groq | groq | GROQ_API_KEY | llama-3.3-70b-versatile | |
| Google | google | GOOGLE_GENERATIVE_AI_API_KEY | gemini-2.0-flash | |
| DeepSeek | deepseek | DEEPSEEK_API_KEY | deepseek-chat | |
| Azure OpenAI | azure | AZURE_OPENAI_API_KEY | gpt-4o-mini | requires attackerLlm.baseURL |
| OpenAI-compatible | openai-compatible | OPFOR_API_KEY | (no default) | requires attackerLlm.baseURL |
openai-compatible covers LiteLLM, OpenRouter, Ollama, vLLM, and any other OpenAI-shaped endpoint — point baseURL at it.
Configuring a model
"attackerLlm": {
"provider": "openai",
"model": "gpt-4o-mini",
"apiKeyEnv": "OPENAI_API_KEY"
},
"judgeLlm": {
"provider": "anthropic",
"model": "claude-3-5-haiku-20241022",
"apiKeyEnv": "ANTHROPIC_API_KEY"
}
judgeLlm is optional — it falls back to attackerLlm when omitted. For azure and openai-compatible, add "baseURL".
Setting keys
export OPENAI_API_KEY=sk-...
export GROQ_API_KEY=gsk_...
export ANTHROPIC_API_KEY=sk-ant-...
.env from the working directory automatically. Load a non-default path with --env <path>.
Add .env to .gitignore. The config stores the env var name (apiKeyEnv), never the key value.
