Introducing the Chatbot Guardrails Arena
Hugging Face Blog
Lighthouz AI and Hugging Face launched the Chatbot Guardrails Arena, a community stress-testing platform where users attempt to trick AI chatbots into revealing sensitive financial data like SSNs, account numbers, and balances. The arena evaluates 12 distinct guardrailed language models—including GPT-3.5, Gemini Pro, Llama-2, and Mixtral—combined with safety frameworks like NVIDIA's NeMo Guardrails and Meta's LlamaGuard, ranked via an Elo rating system based on community votes. Results will be published on a public leaderboard to identify which models and guardrails best preserve privacy, informing development of more secure AI assistants and agents.
