RSAIRF
Open Ethics Initiative logo
AI RISK GOVERNANCE

Really Simple
AI Risk Framework

AI risk governance framework with a shared register for identifying, scoring, and tracking what can go wrong when you build, ship, or run AI — thirty risks, four lifecycle stages, one ID space.

RSAIRF · Open Ethics Initiative
github.com/OpenEthicsAI/RSAIRF
Licensed CC BY 4.0
01–30

Why it exists

Organizations deploy and use AI. Few build it. Almost none monitor it well.

The AI Risk Register is curated from public sources and the internal incident learnings of real client engagements — not written from theory. It exists because the biggest exposures usually aren’t in the model itself. They’re in the gaps around it: tools adopted without approval, and monitoring that was never built.

AIR 01–30
30 traceable IDs
4
four lifecycle stages
5
mapped to five existing governance models

The lifecycle

Most initiatives move through four stages

Development
Model Development
Deployment
Model / Component Deployment
Use
AI Use
Continuous Monitoring & Evaluation — runs across all three stages, and feeds findings back into any of them
Development
You design or materially change AI behavior — data selection, training, tuning, evaluation.
Deployment
You don’t build the model, but you integrate, configure, and release AI components into production.
Use
You consume a deployed AI solution and make decisions or take actions based on its outputs.
Monitoring
You track risk, quality, and incidents over time — continuous, and running across every other stage.

The process

Four steps, repeated at every transition

Check which lifecycle stages apply to your initiative — run a discovery session with your technical team if it’s unclear.

Map risks using register.csv. Stage tags show where each risk is typically introduced or first exploitable.

For every applicable AIR## risk, record likelihood, impact, an owner, and a mitigation — and reference that ID in your project’s risk log or audit.

Found a risk that isn’t represented? Open a repository issue — your contribution may help other teams too.

Re-check the register at each stage transition — a risk that was out of scope earlier may now apply.
↻ treat this as a living list, not a one-time checklist

Scoring · Likelihood

How probable it is

For every applicable risk, rate likelihood — how probable it is that the risk materializes in your context.

ScoreLikelihoodMeaning
1RareNot expected during the initiative’s lifetime.
2UnlikelyCould occur, but not expected.
3PossibleMay occur occasionally.
4LikelyExpected to occur at some point.
5Almost certainExpected to occur often, or already observed.

Scoring · Impact

How severe, if it happens

Then rate impact — the severity of the consequence if the risk does materialize.

ScoreImpactMeaning
1NegligibleMinimal effect; no material harm.
2MinorLimited harm, easily remediated.
3ModerateNoticeable harm, cost, or user impact.
4MajorSignificant financial, regulatory, or reputational harm.
5SevereCritical or potentially irreversible harm to people, rights, or the organization.
Risk score = Likelihood × Impact  ·  range 1–25

The matrix

Likelihood × impact, at a glance

The same 1–25 score, banded into four priority levels — consistent across initiatives, without heavy tooling.

Impact →
1
2
3
4
5
1 · Rare
1
2
3
4
5
2 · Unlikely
2
4
6
8
10
3 · Possible
3
6
9
12
15
4 · Likely
4
8
12
16
20
5 · Almost certain
5
10
15
20
25
Likelihood ↓
Priority band
Low · score 1–4
Medium · score 5–9
High · score 10–14
Critical · score 15–25

The register

Thirty risks, mapped across four stages

Use
23
Development
13
Deployment
12
Monitoring
6

Counts sum to more than 30 because several risks are tagged to more than one lifecycle stage.

Spotlight

A sample from the register

AIR01

Shadow AI Deployments

AI tools or models deployed and used by internal teams without documentation, approval, or oversight from security or governance.

DEPLOYMENT · USE
AIR03

Data Leakage

Sensitive data exposed through third-party tools, API calls, or model interactions, leading to loss of confidentiality or regulatory breach.

DEVELOPMENT · USE
AIR15

Adversarial Attacks

An attacker deliberately alters input data to mislead the model — from prompt injection to imperceptible perturbations.

USE
AIR17

Bias & Discrimination in Outputs

Skewed or discriminatory results from biased training data or model design, including unequal performance across demographic groups.

DEVELOPMENT · USE
AIR27

Unbounded Consumption

Uncontrolled or excessive inference requests degrade availability or generate runaway compute costs — a.k.a. Denial of Wallet/Service.

USE · MONITORING
AIR29

Multi-Agent & Agentic AI Risks

Risks specific to autonomous agents acting with tools and memory via an agent harness — unconstrained tool access, unsafe actions, and privilege escalation.

DEVELOPMENT · DEPLOYMENT · USE

Full register

All thirty, AIR01–AIR30

IDRiskStage
AIR01Shadow AI DeploymentsDeployment, Use
AIR02Supply Chain VulnerabilitiesDevelopment, Deployment
AIR03Data LeakageDevelopment, Use
AIR04Cross-Border Data TransferDeployment, Use
AIR05No Validation / Improper OutputUse
AIR06Insecure Deployment PipelinesDeployment
AIR07Critical 3rd-party DependencyDevelopment, Deployment
AIR08Excessive Agency / Over-Reliance on AIUse
AIR09Model PoisoningDevelopment
AIR10Model Inversion / Membership Inference / TheftDeployment, Use
AIR11Poor MonitoringMonitoring
AIR12Feedback Loop ContaminationUse, Monitoring
AIR13Alert FatigueMonitoring
AIR14Insider ThreatsDevelopment, Deployment, Use
AIR15Adversarial AttacksUse
IDRiskStage
AIR16Model DriftUse, Monitoring
AIR17Bias & Discrimination in OutputsDevelopment, Use
AIR18Critical Unintended ConsequencesDevelopment, Deployment, Use
AIR19Censorship / Guardrail interferenceUse
AIR20Deployment misfitDeployment, Use
AIR21Lack of Explainability / Provenance opacityDevelopment, Use
AIR22No AccountabilityDevelopment, Deployment, Monitoring
AIR23Unclear Model Ownership / IPDevelopment, Use
AIR24Output Integrity TamperingUse
AIR25System Prompt / Instruction LeakageUse
AIR26RAG / Vector Store VulnerabilitiesDevelopment, Use
AIR27Unbounded ConsumptionUse, Monitoring
AIR28Malicious Use / Out-of-Scope ExploitationDeployment, Use
AIR29Multi-Agent & Agentic AI RisksDevelopment, Deployment, Use
AIR30Environmental / Sustainability ImpactDevelopment, Use

Standards

Mapped to existing governance models

Every AIR## risk is cross-referenced so findings plug straight into the programs you already run.

NIST AI RMF
Govern · Map · Measure · Manage
OWASP LLM Top 10
2025 edition
EU AI Act
Articles 5–95
ISO/IEC 42001
Annex A controls
MITRE ATLAS
AML.T#### techniques
AIRRiskOWASP LLMEU AI ActMITRE ATLAS
AIR03Data LeakageLLM02 Sensitive InfoArt 10AML.T0057
AIR15Adversarial AttacksLLM01 Prompt InjectionArt 15AML.T0051
AIR26RAG / Vector StoreLLM08 Vector WeaknessArt 10 & 15AML.T0070
AIR27Unbounded ConsumptionLLM10 UnboundedArt 15AML.T0029

Full mapping in standards-map.csv.

In practice

From register to risk log

Record two ratings, an owner, and a control per applicable risk in your project’s log — traceable back to the register by AIR ID.

InitiativeAIRRiskLIScoreLevelOwnerStatus
Support chatbotAIR05No Validation / Improper Output3412HighML LeadMitigating
Support chatbotAIR27Unbounded Consumption236MediumPlatformAccepted

Template fields: likelihood, impact, score, level, controls, owner, sign-off, KRI / monitoring signal, next review, status — see risk-log.template.csv.

What’s inside

Four files, ready to use

register.csv

The 30-risk AI Risk Register with IDs, lifecycle tags, and descriptions.

standards-map.csv

Cross-references to NIST, OWASP, EU AI Act, ISO 42001 & MITRE ATLAS.

risk-log.template.csv

Ready-to-fill log for scoring, ownership, sign-off, and review cadence.

README.md

Lifecycle model, usage workflow, and scoring guidance.

Open Ethics Initiative logo

Get involved

Treat it as a living framework.

If you identify a risk not represented here, propose it by opening an issue in the repository — your contribution may help other teams facing the same gap.

License
CC BY 4.0
Contact
contact@openethics.ai