An Introduction to AI Secure LLM Safety Leaderboard

Hugging Face just rolled out a new LLM Safety Leaderboard, and it's built on top of DecodingTrust, a research project from the Secure Learning Lab that picked up an Outstanding Paper Award at NeurIPS 2023. The idea is simple to state and hard to execute: before companies keep shipping language models into products, someone needs a rigorous way to check how those models behave when pushed, prodded, and deliberately misled.

DecodingTrust doesn't just run a single toxicity check and call it a day. It grades models across eight separate angles — toxicity, stereotype bias, adversarial robustness, out-of-distribution robustness, resistance to misleading demonstrations, privacy leakage, machine ethics, and fairness. Each category comes with its own attack playbook. For toxicity, the team built 33 tricky system prompts, things like asking a model to role-play or respond as if it were code, then scored the outputs with Google's Perspective API. For bias, they tested 24 demographic groups against 16 stereotype topics, repeating each prompt five times to average out noise. Privacy testing went as far as coaxing models to spit out email addresses and credit card numbers using different phrasings, which is where things get genuinely unsettling.

One finding stands out from the paper: GPT-4, widely assumed to be the safer, smarter sibling of GPT-3.5, actually turned out to be more vulnerable in several of these tests. That's not a small footnote — it undercuts the easy assumption that newer or bigger automatically means safer. The researchers also found that models understand privacy-related language inconsistently. Ask GPT-4 to keep something