Research
Welcome to the 0DIN Research Hub, here we delve into the methodologies, taxonomies, and scoring systems that underpin our understanding of AI risks and mitigations.
How we measure AI safety
Three steps from attack to rating.
What's in Scope
Eligible models & security boundaries
Every evaluation starts with a clear scope. Which models are in play and which security boundaries define a meaningful failure. Check which models and boundaries are eligible for rewards.
How We Score Each Threat
Five standardized tests,
one repeatable score
Our Jailbreak Evaluation Framework (JEF) quantifies jailbreaks by how many models they break, how flexibly they retarget across subjects, and yielding a transparent severity score from 0 to 10.
See the Results
Live security rankings across LLMs
Live security profiles for every major LLM: vulnerability distributions, Attack Success Rate (ASR), and trend data updated continuously.
Submit Your First Vulnerability
For ResearchersEverything you need to find, validate, and submit a bug bounty report.
Additional References
Jailbreak Taxonomy
A structured hierarchy of bypass techniques
0DIN's Jailbreak Taxonomy organizes bypass techniques into a structured hierarchy, giving researchers a shared vocabulary to describe and compare attacks.
Social Impact Score (SIS)
How severity and rewards are calculated
Rewards are tied to real-world harm potential, not technical novelty. SIS weighs impact from minor policy violations to life-threatening content.
Nude Imagery Rating System (NIRS)
Standardized severity scale for explicit imagery
NIRS provides a consistent way to rate the severity of explicit imagery produced by image models, so submissions can be triaged and rewarded objectively.