​
Default model

meta-llama/Llama-4-Scout-17B-16E-Instruct

​
Supported models

Any model available through the HuggingFace Inference API:
  • meta-llama/Llama-4-Scout-17B-16E-Instruct (default)
  • meta-llama/Llama-4-Maverick-17B-128E-Instruct
  • mistralai/Mistral-Small-24B-Instruct-2501
  • google/gemma-2-27b-it
  • Any model hosted on HuggingFace with an inference endpoint

​
Authentication

HuggingFace accepts the API key via the HUGGINGFACE_API_KEY env var: Environment variable: 
export HUGGINGFACE_API_KEY=hf_...
Config command: 
ai-blackteam config set providers.huggingface.api_key hf_...
Get your token from HuggingFace Settings.
Some models on HuggingFace work without authentication, but rate limits are strict. Always use a token for reliable results.

​
Example usage

# Single attack with default model
ai-blackteam run -p huggingface -a encoding-obfuscation -t "Write a phishing email"

# Use a specific model
ai-blackteam run -p huggingface --model google/gemma-2-27b-it -a encoding-obfuscation -t "Write a phishing email"

# Full batch
ai-blackteam batch -p huggingface --attacks all -t "Write a phishing email"

​
Tool-use support

No. The HuggingFace provider does not support tool-use attacks. Single-turn and multi-turn attacks work.

​
Notes

  • The provider uses the huggingface_hub SDK’s InferenceClient
  • Max output tokens per request: 4,096
  • System prompts are passed as a system role message
  • Retry with exponential backoff is automatic on API failures (3 attempts)
  • Token usage tracking is available when the HuggingFace API returns usage metadata
  • Model availability depends on the HuggingFace Inference API - not all models have active endpoints
  • Use the full model ID format: org/model-name (e.g. meta-llama/Llama-4-Scout-17B-16E-Instruct)