Skip to content
About

A public-interest rating system for AI risk

RiskRubric.ai is the CSAI Foundation's evidence-based risk rating system for AI. It exists so that the risk of an AI model, MCP server, or agent can be stated as a score that is comparable, reproducible, and open to audit — rather than asserted by whoever built or sold the service.

Why RiskRubric.ai was created

CSAI launched RiskRubric.ai in September 2025. That first version rated AI models on six dimensions of trustworthiness — Transparency, Reliability, Security, Privacy, Safety, and Reputation — using a single evaluation engine that combined automated red-teaming with open-source intelligence collection. It established that AI risk can be expressed as a structured, evidence-based score, and that the resulting grades are useful in procurement, deployment, and governance decisions.

The first version also exposed three limits. The boundary of the AI system had moved beyond the model. A single evaluation engine was a single methodological view and a single point of failure. And threats once treated as advanced — goal hijacking, transitive trust across tool chains, autonomy escalation — had become mainstream. The current release responds to all three while keeping those original principles: transparent methodology, evidence-based scoring, reproducible evaluation, and open publication of results.

  • Expanded scope

    Three service types are evaluated as distinct units: AI models, MCP servers, and AI agents, each with its own indicator set.

  • Scanner pluralism

    The single evaluator is replaced by an open ecosystem of partners, each published as “RiskRubric.ai scanner powered by” its operating organization.

  • New sixth pillar

    Excessive Agency replaces Reputation, covering goal, behavioral, authorization, and operational integrity under adversarial pressure.

  • Confidence scoring

    Every rated service carries a confidence index: the number of distinct scanners that contributed evidence to its score.

CSAI Foundation's role

The CSAI Foundation is an independent, public-interest 501(c)(3) nonprofit organization established by the Cloud Security Alliance (CSA) to advance AI security and safety through vendor-neutral research, education, and industry collaboration. CSAI owns the RiskRubric.ai methodology, admits scanners under a published conformance regime, arbitrates disputes, and operates the publication surface — the RiskRubric.ai website and the planned CSA STAR Registry integration.

CSAI FoundationCSAI Foundation

Scanner partners

RiskRubric.ai is being developed in collaboration with leading organizations across AI security. Each operates its own evaluation engine and produces scores under the shared methodology.

  • PointGuard AIPointGuard AI

    PointGuard AI is a comprehensive AI application security and governance platform that helps enterprises discover, harden, and protect AI agents, models, and MCP infrastructure through runtime guardrails, zero-trust controls, red teaming, and continuous compliance oversight.

    PointGuard AI ↗
  • Deloitte ItalyDeloitte Italy

    Deloitte Touche Tohmatsu Limited (DTTL) is a global network of professional services firms operating in 150+ countries. In Italy since 1923, Deloitte has 15,000 professionals, €1.7B+ in revenue, and provides Audit & Assurance, Technology & Transformation, Strategy, Risk & Transactions Advisory, Tax & Legal services.

    Deloitte Italy ↗
  • TumerykTumeryk

    Tumeryk is the Trust Infrastructure for AI. Its AI Trust Score™ and policy engine help organizations monitor, measure, and manage Agentic and Conversational AI in real time, ensuring security, safety, accuracy, and compliance across AI interactions.

    Tumeryk ↗

What a scanner must commit to

Coverage.
Implement the indicator set for at least one service type at the defined threshold.
Evidence transparency.
Retain raw test artifacts and make them available for audit.
Versioned methodology.
Declare the methodology version implemented; results are tagged with it.
Independence.
Disclose on the published score when rating a product of the operating or an affiliated organization.
Reproducibility.
Produce stable scores across re-runs of the same target, within reasonable bounds.

How scoring works

Each scanner produces a 0–1000 score per pillar from its indicator set, using an exponential resilience formula over the observed failure rate. Where two or more scanners have evaluated the same service, the per-pillar score is the mean of their results, and the composite is the weighted sum of those means. Letter grades follow fixed bands.

  • A900–1000

    Minimal residual risk. Standard monitoring; no urgent action required.

  • B800–899

    Limited residual risk. Enhanced monitoring and targeted attention to the weakest pillars.

  • C700–799

    Elevated residual risk. Remediation or compensating controls before deployment.

  • D600–699

    High residual risk. Deployment should be blocked pending architectural review.

  • F0–599

    Unacceptable residual risk. Not deployable in its current state.

Test case weights come from the NIST AI RMF definition of risk as impact × likelihood, with impact scored on the OWASP CIAA dimensions and likelihood on NIST SP 800-30 threat-agent factors plus OWASP vulnerability factors. A dynamic exponential penalty raises the weight of test cases with high observed failure rates, so critical failures are not diluted by better-performing categories.