Skip to content
Evidence-Based AI Risk Ratings from CSAI

Know the risk before you deploy.

Each rated AI model and MCP server carries a risk score, a letter grade, and a breakdown across its risk pillars, so you can see where a service is strong and how it stands against others.

risk pillars
6
independent scanner partners
3
score, mapped to A–F
0–1000

The problem: vendor claims are not evidence

Teams choosing an AI model or wiring up an MCP server have to rely on marketing material, anecdotal reports, or their own ad-hoc testing. There is no shared, comparable measure of how a service behaves under adversarial pressure — so procurement, deployment, and governance decisions get made without it.

RiskRubric.ai translates technical findings into scores, grades, and supporting indicators, so that the same question can be asked of every service and answered the same way.

  • Security and governance leads

    Validate that a model meets a risk threshold before it reaches production, and keep a record of why.

  • Developers and architects

    Check a scorecard instead of defaulting to whatever model is most popular or most powerful.

  • Procurement and executives

    Compare candidate services on a standardized basis, with the evidence behind each score retained for audit.

Why now

  • The AI system boundary has moved

    Production deployments now include MCP servers exposing tools and data, and agents composing models, tools, and memory under autonomous control. Assessment confined to the model layer leaves substantial attack surface unmeasured.

  • One evaluator is one point of failure

    A single evaluation engine represents a single methodological view and a constraint on coverage. Cross-evaluator triangulation reduces systematic blind spots and scales assessment volume.

  • The threat model has matured

    Goal hijacking, transitive trust across tool chains, and autonomy escalation were once treated as advanced concerns. They are now mainstream and need first-class treatment in any credible rating framework.

How it works

CSAI governs the methodology. Independent partner organizations operate the evaluation engines — called scanners — that produce the evidence.

  1. 1

    Scanners test the service

    Each scanner runs the RiskRubric.ai indicator set against the target — adversarial red-teaming plus open-source intelligence — and retains the raw test artifacts for audit.

  2. 2

    Results roll up to pillar scores

    Indicators aggregate into a 0–1000 score per pillar. Test cases that fail more often carry more weight, so a critical failure is not smoothed away by average performance elsewhere.

  3. 3

    Scores are published with confidence

    Pillar scores combine into a composite score and A–F grade. A confidence index records how many independent scanners contributed evidence, and per-scanner results are shown side by side.

The six pillars

Weights are fixed by the current methodology for each service type.

  • —TransparencyHow clearly a service discloses its nature, capabilities, limitations, and operational behavior, and how observable its decisions and tool interactions are to users and auditors.
  • —ReliabilityConsistency, completeness, and accuracy of outputs and actions across repeated or varied conditions, including recovery from partial failure.
  • —SecurityResilience to adversarial attack, misuse, and evasion, and the integrity of the autonomy controls that bound a service's actions to its intended scope.
  • —PrivacyHandling of personal and sensitive data, including unauthorized collection, retention, exposure, or solicitation across tasks, tools, and tenants.
  • —Safety & Societal ImpactsAvoidance of harmful, manipulative, or ethically problematic outputs and behaviors, including bias, misinformation, harassment, and illegal-activity assistance.
  • —Excessive AgencyGoal, behavioral, authorization, and operational integrity: whether the service stays inside its scope and boundaries under adversarial pressure.

The current framework replaced the earlier Reputation pillar with Excessive Agency: Reputation carried 5% weight, conflated vendor track record with model behavior, and could not be independently evidenced by scanners.

Browse models that have been scored

The highest and lowest composite scores currently published, to show the spread you are choosing within.

View the full rankings →

Loading rankings…

Recently scored services

Browse all services →

Loading services…

Who is behind RiskRubric.ai

The CSAI Foundation, a public-interest 501(c)(3), owns the RiskRubric.ai methodology and operates the publication surface. It does not sell evaluations. Scanners are commercial, independent, and operate inside a published conformance regime they do not control — including disclosure when a scanner rates a product of its own organization.

  • PointGuard AIPointGuard AI
    PointGuard AI
    AI application security and governance platform
  • Deloitte ItalyDeloitte Italy
    Deloitte Italy
    Audit, assurance, and risk advisory network
  • TumerykTumeryk
    Tumeryk
    Trust infrastructure and policy engine for AI