# Democratizing AI Safety with RiskRubric.ai

[Hugging Face](https://yomu.fyi/company/hugging-face) · Gal Moyal · Sep 18, 2025

**Type:** Announcement

## Summary

Cloud Security Alliance and Noma Security introduced RiskRubric.ai to provide standardized, transparent risk assessments across the open AI model ecosystem. The framework evaluates AI models across six pillars—transparency, reliability, security, privacy, safety, and reputation—using over 1,000 reliability tests, 200 adversarial security probes, automated code scanning, and harmful content evaluations. Each model receives 0–100 scores and A–F letter grades, supplemented by specific vulnerability findings and recommended mitigation strategies to assist deployment filtering. Initial benchmark results across models showed composite scores ranging from 47 to 94 with a median of 81, revealing polarized safety distributions and indicating that security hardening directly correlates with reduced safety risks.

## Context

With more than 500,000 models available on the Hugging Face hub, developers lack a systematic, standardized method to evaluate model security posture, privacy implications, and potential failure modes prior to deployment.

## Approach / What changed

Cloud Security Alliance and Noma Security launched RiskRubric.ai to evaluate models across six pillars: transparency, reliability, security, privacy, safety, and reputation. The platform automates evaluations using over 1,000 reliability tests, 200 adversarial security probes for jailbreaks and prompt injections, automated code scanning, documentation reviews, privacy assessments, and structured harmful content tests. These produce 0-100 scores and A-F letter grades alongside vulnerability reports and remediation guidance.

## Takeaways

- Model risk scores ranged from 47 to 94 with a median of 81, showing polarization where 54 percent reached A or B levels while a long tail clustered in the medium-to-low protection C and D range.
- Safety pillar scores exhibited the widest variation across evaluated models but tracked closely with security posture, indicating that prompt injection defenses and policy enforcement directly mitigate harmful outputs.
- Stricter guardrails frequently decrease user-perceived transparency through opaque refusals, which can be mitigated by combining safeguards with explanatory refusals and provenance signals.

**Tags:** [LLMs](https://yomu.fyi/topic/llm), [Machine Learning](https://yomu.fyi/topic/machine-learning), [Open Source](https://yomu.fyi/topic/open-source), [Privacy](https://yomu.fyi/topic/privacy), [Testing](https://yomu.fyi/topic/testing)

[Read original post](https://huggingface.co/blog/riskrubric)