Tool connections
How the tools on this site fit together. Each connection says what it links, which versions it was built on, and how far to trust it.
AI Risk Map™ × Four Faces of AI Risk™
The AI Risk Map lists the risk factors that come with your own use of AI, organized by who owns them. Four Faces sorts AI risk by cause. This mapping places every active AI Risk Map factor in one or two faces, so you can start from either tool and find the other's view.
The AI Risk Map covers faces 1 to 3, the faces about our own AI. Face 4 is outside its subject boundary by design, so that column holds no factors.
Get the mapping
- Face per factor, IDs onlyJSON
- AI Risk Map with a Four Faces columnCSV
In the AI Risk Map repository, under companions/four-faces/.
Domains by face
Each domain sits in the face where most of its factors fall. A dashed block is the part of a domain whose factors sit in another face, labelled with the categories they come from.
Mapping revision 1 · AI Risk Map v2.1 · Four Faces v1.2
Are we making the right bets?
02Strategy & Value Realization
AIRM-193Our advantage erodes as competitors and customers approximate it with AIAIRM-194AI agents disintermediate us from our customersAIRM-026Failure to weigh the competitive risk of not adopting AI (inaction)AIRM-020Lack of a coherent AI strategy aligned to objectives and risk appetiteAIRM-021AI adoption driven by hype rather than demonstrated business valueAIRM-022Inability to adapt the roadmap as foundation-model capabilities shiftAIRM-204Fear-driven AI bans push use into the shadowsAIRM-023Use cases chosen without a clear problem, success measure, or risk–benefit weighingAIRM-024‘Pilot purgatory’ — pilots stalling on data readiness, evaluation difficulty, or inability to demonstrate valueAIRM-025Misjudged AI sourcing fork: train, fine-tune, RAG, or APIAIRM-027Unrealistic ROI expectations, including returns that need workflow redesignAIRM-028Inability to attribute value actually delivered by AI investmentsAIRM-029Hidden total cost of ownership (data, talent, integration) eroding returns
10Workforce, Human & Societal
AIRM-151Shortage of AI and AI-risk talent, competing against AI-native employers for scarce skillsAIRM-152Inability to retain specialists, who leave taking undocumented model and prompt knowledge with themAIRM-195Leaders manage AI like a conventional IT rolloutAIRM-154Erosion of staff expertise through over-delegation to AIAIRM-155Loss of institutional knowledge as AI replaces human processesAIRM-157Change management underestimated as AI reshapes the work itselfAIRM-158Workforce resistance or low adoption driven by displacement fear or distrust of AI outputAIRM-159Labor-relations and morale impacts of AI deploymentAIRM-161Customers distrust or avoid AI-mediated services
11Financial & Resource
AIRM-165Poor AI cost forecasting as usage scales to productionAIRM-166AI's energy and water use unmeasured and unweighedAIRM-167Stranded investment, including AI scaled too early or funded past failureAIRM-168Procurement exposure where providers reprice tokens, tiers, or context windows unilaterallyAIRM-170Compute, GPU, or power scarcity constraining scalingAIRM-172Concentration of compute supply among few providers
01Governance (part)Governance & oversight · Roles & accountability · Risk-management integration
AIRM-001Inadequate executive/board oversight of AI adoption and riskAIRM-003Absence of an AI governance body or charter with a clear mandateAIRM-004Unclear ownership and accountability (RACI) for AI risk decisionsAIRM-005No clear executive owner for AI adoptionAIRM-012AI risk not integrated into enterprise risk management (ERM)AIRM-014No defined risk appetite or tolerance thresholds for AI
09Third-Party (part)Foundation-model dependence · Vendor due diligence · Lock-in & portability
AIRM-139Concentration risk from dependence on a few model providersAIRM-140Model monoculture, including one model's blind spot reproduced across everything it producesAIRM-141Provider deprecation or retirement of a model forcing unplanned migrationAIRM-143Vendor viability risk in an unusually volatile provider market with frequent pivots and failuresAIRM-145Vendor / model lock-in limiting portability and leverageAIRM-146High switching cost from proprietary APIs and fine-tuningAIRM-147Lack of exit provisions, where fine-tunes, embeddings, and prompt assets do not transfer between providers
Is it well built and well run?
01AI Governance & Accountability
AIRM-001Inadequate executive/board oversight of AI adoption and riskAIRM-002Ungoverned ‘shadow AI’ adoption outside sanctioned channelsAIRM-003Absence of an AI governance body or charter with a clear mandateAIRM-004Unclear ownership and accountability (RACI) for AI risk decisionsAIRM-005No clear executive owner for AI adoptionAIRM-006Accountability for agent decisions undefined, internally and across the AI stackAIRM-007Segregation of duties collapses across AI config, operation, and reviewAIRM-008Absence of an AI acceptable-use policy and model-approval gatesAIRM-009AI policies not kept current with evolving capabilities and regulationAIRM-010Inconsistent standards as AI proliferates faster than governanceAIRM-011No agreed definition of ‘AI’, leaving policy and inventory scope undefinedAIRM-012AI risk not integrated into enterprise risk management (ERM)AIRM-013Failure to conduct AI impact assessments for individuals and societyAIRM-014No defined risk appetite or tolerance thresholds for AIAIRM-015Inability to audit AI: no complete inventory of models, agents, and connectorsAIRM-016Insufficient internal-audit competence to assess AI controlsAIRM-017Absence of independent assurance over AI control effectivenessAIRM-018AI output relied on in financial-reporting controls without scope changeAIRM-019No way to identify which artifacts AI produced, so a model flaw cannot be scoped
03Data & Privacy
AIRM-030Unverified provenance or integrity of training, fine-tuning, or RAG dataAIRM-031Reliance on data scraped without lawful basis or clear rightsAIRM-032Inadequate data lineage to trace data through the AI pipelineAIRM-033Poor data quality undermining model reliabilityAIRM-034Inadequate or biased labeling introducing systematic errorAIRM-035Stale data, including grounding and retrieval sources that drift out of date behind a confident-sounding modelAIRM-036Personal data used for training without consent or purpose limitationAIRM-037Inability to honor data-subject rights once data is in a modelAIRM-038Re-identification of individuals from aggregated or model-derived dataAIRM-039No guardrails keeping confidential or personal data out of model outputAIRM-040Unclassified sensitive data in AI, so handling controls never applyAIRM-041Confidential data pasted into an approved AI tool, leaving despite working controlsAIRM-042Use of copyrighted or licensed material in training without rightsAIRM-043Contamination of datasets with third-party IP or trade secretsAIRM-044Inability to prove you held rights to the training data usedAIRM-045Failure to preserve prompts, outputs, and agent traces subject to litigation hold or e-discoveryAIRM-046Deletion at source does not remove data absorbed into modelsAIRM-047Persistence of sensitive data in embeddings, caches, or model weightsAIRM-048Failure to remove data across distributed / replicated AI stores
04Model Development & Performance
AIRM-049Inappropriate model/algorithm choice for the problem and risk contextAIRM-050Poorly documented or untracked model developmentAIRM-051Insecure or unmanaged hyperparameters and training configurationAIRM-052Hallucinated or meaning-changing output accepted as reliableAIRM-053Validation data unrepresentative; real performance shows only with live usersAIRM-054Skipped or insufficient red-teaming, safety, and fairness testingAIRM-055Absence of independent model validation / effective challengeAIRM-056Model drift degrading performance as data / conditions changeAIRM-057Lack of robustness to edge cases and distribution shiftAIRM-058Small changes to the input produce very different outputAIRM-059Black-box models used in high-stakes decisions without interpretabilityAIRM-060Inability to produce adverse-action / decision explanationsAIRM-061Over-reliance on post-hoc explanations that misrepresent the modelAIRM-062Non-reproducible training pipelines undermining validation and auditAIRM-063Inadequate model documentation / model cardsAIRM-064Loss of traceability between model versions, data, and results
06Safety, Ethics & Responsible AI
AIRM-086Discriminatory or biased outputs causing disparate impactAIRM-087Inequitable performance across languages, demographics, and disabilitiesAIRM-088Bias amplification / homogenization from feedback and scaleAIRM-089Culturally inappropriate output as AI scales across regionsAIRM-090Generation of toxic, dangerous, or hateful contentAIRM-091Harmful, misleading, or defamatory output driving decisionsAIRM-092Inappropriate or unsafe AI output reaching minors or other vulnerable usersAIRM-199No guardrails on what customer-facing AI says in our nameAIRM-202AI makes false statements about real people or organizationsAIRM-093Failure to disclose to users that they are interacting with AIAIRM-094Undisclosed synthetic media / deepfakes lacking provenanceAIRM-095Inadequate transparency to affected parties about AI's roleAIRM-192Our AI service changes under customers without notice, so their validations go staleAIRM-098AI optimized for a proxy measure that diverges from users' interestsAIRM-099User over-trust / automation bias in AI outputsAIRM-100Anthropomorphism and emotional dependence on conversational AIAIRM-101Erosion of meaningful human control over decisionsAIRM-102Review capacity cannot scale to AI output volume, degrading oversight to rubber-stamping
07Operations & Resilience
AIRM-103Immature MLOps / fragile deployment pipelinesAIRM-104Inconsistent or unsanctioned (shadow) model deploymentAIRM-105Environment and config management gaps undermining reproducibilityAIRM-106Training-serving skew degrading live performanceAIRM-107AI tool-execution or test containment fails and AI reaches real systemsAIRM-108No monitoring, acted on, for drift, biased outcomes, or abuseAIRM-109Inadequate logging of inputs, outputs, tool calls, agent decisions, and who authorized themAIRM-110Non-determinism making anomalous behavior hard to define or detectAIRM-111Sanctioned agent activity indistinguishable from agentic attack in telemetryAIRM-112Inference latency or unavailability degrading dependent servicesAIRM-113Inability to scale inference to peak demandAIRM-115Immature AI incident response and escalation processesAIRM-116Inability to investigate AI incidents due to non-determinism / poor logsAIRM-117Inadequate content-moderation / trust-&-safety operations at scaleAIRM-118Defensive model refuses to analyze attack data mid-incidentAIRM-201No tested off-switch or fallback for a misbehaving AI featureAIRM-119Vendor or model updates changing behavior, often silently or without noticeAIRM-120Inadequate versioning, change control, and tested rollback to a known-good model versionAIRM-121Unmanaged model decommissioning / sunset leaving orphaned dependenciesAIRM-122Uncontrolled continuous learning introducing driftAIRM-123Controls re-verified less often than the AI system changesAIRM-124Prompts and retrieval config changed outside change control
08Legal, Regulatory & Compliance
AIRM-125Unknowingly deploying a prohibited or high-risk AI use caseAIRM-126AI obligations not identified or built in before deploymentAIRM-127Inability to keep pace with fragmented, fast-changing AI regulationAIRM-128User inability to contest or seek recourse on an AI-driven decisionAIRM-129Misclassified as deployer when fine-tuning makes you a providerAIRM-200No way to find and make whole everyone a flawed AI decision affectedAIRM-130Unsettled liability for harm caused by AI decisionsAIRM-131AI-generated output that closely reproduces or imitates others' copyrighted work or trademarksAIRM-132Uncertain ownership / copyrightability of AI-generated contentAIRM-133Vendor contracts lacking AI terms: training use, output ownership, change noticeAIRM-134Use of open-weight or licensed models outside their license termsAIRM-135Failure to update contracts as AI use and regulation evolveAIRM-136Data crossing borders through AI services, including hidden routingAIRM-137Conflicting regional AI requirements the organization cannot reconcile for a single system (externally imposed)AIRM-138Inability to evidence and audit AI compliance across jurisdictions (internal evidence burden)
09Third-Party & Model Supply Chain
AIRM-140Model monoculture, including one model's blind spot reproduced across everything it producesAIRM-142Vendor diligence missing: are prompts retained, trained on, or shared?AIRM-188No way to verify a vendor's 'AI-powered' claim before buying itAIRM-144Accountability for a vendor's AI failures staying with the organization, whatever the contract saysAIRM-149Open-source AI dependencies abandoned by their maintainersAIRM-150Opaque sub-processors, including undisclosed models and embedded AI featuresAIRM-189Models deployed without verifying integrity or lineageAIRM-191Safety training stripped or eroded in a self-hosted model
12Autonomy & Delegated Action (Agentic AI)
AIRM-173Agent pursues unintended goals or misgeneralizes its objectiveAIRM-174Agent pursues an authorized objective through unauthorized meansAIRM-175Excessive autonomy / agency beyond intended mandateAIRM-176No working interrupt or kill switch for an agent mid-taskAIRM-206No human approval before high-consequence agent actionsAIRM-203Agents that rewrite their own instructions or toolsAIRM-177Agent granted a broader set of invokable tools / functions than its task requiresAIRM-179Third-party tool/connector servers as an unvetted execution surfaceAIRM-182Persistence of corrupted or sensitive data in agent memoryAIRM-183Cascading failures or loops across interacting agentsAIRM-184Misplaced inter-agent trust or agent collusionAIRM-185Emergent, unpredictable behavior in multi-agent systemsAIRM-186Agents able to take irreversible real-world actionsAIRM-187Accountability gap for harm caused by autonomous actionAIRM-190Third parties can't tell your agent acted, so no one can warn you
05Security (part)Output & integration security · AI access & identity
AIRM-075Unsanitized output reaching downstream interpretersAIRM-077Over-trusted output triggering unsafe automated actionsAIRM-085Retrieval serving content beyond the user’s own entitlements
10Workforce (part)Skills & talent · Deskilling & loss of institutional knowledge · Societal & reputational
AIRM-152Inability to retain specialists, who leave taking undocumented model and prompt knowledge with themAIRM-153Inadequate AI literacy across the workforceAIRM-156Capability atrophy — inability to operate if AI is unavailableAIRM-160Reputational damage amplified when AI fails in publicAIRM-162AI decisions affecting people who never chose to use the systemAIRM-163AI-washing: overstating capability to customers, investors, or regulators
11Financial (part)Cost & compute economics · Investment / procurement
AIRM-164Uncontrolled AI consumption driving up inference spendAIRM-166AI's energy and water use unmeasured and unweighedAIRM-169Insurance excludes AI losses, or defines 'user' as human only
Can it be turned against us?
05Security & Adversarial Threats
AIRM-065Training-data poisoning that degrades or backdoors model behaviorAIRM-066Tampering with RAG / knowledge-base sources consumed at inferenceAIRM-067Label-flipping or adversarial dataset partitioningAIRM-068Model, reasoning, or insight extraction through useAIRM-069Membership-inference or model-inversion exposing training dataAIRM-070Backdoored / trojaned pre-trained model from a public hubAIRM-071Direct prompt injection altering behavior or bypassing guardrailsAIRM-072Indirect / cross-domain prompt injection via retrieved contentAIRM-073Jailbreaks defeating safety controls and restrictionsAIRM-074System-prompt leakage exposing operational instructionsAIRM-197Inputs crafted to evade AI detection modelsAIRM-205Untrusted content treated as instructionsAIRM-075Unsanitized output reaching downstream interpretersAIRM-076Interception or manipulation of the inference output streamAIRM-077Over-trusted output triggering unsafe automated actionsAIRM-078Compromised or malicious model from the AI supply chainAIRM-079Known vulnerabilities in ML dependencies not patched in timeAIRM-080Tampering with model artifacts, prompts, or configurationAIRM-081Model-invented dependencies registered by attackersAIRM-082Weak access controls and usage limits on AI endpointsAIRM-083Over-privileged service accounts and keys for AI componentsAIRM-084Credential sprawl as AI tools and agent frameworks outpace governanceAIRM-085Retrieval serving content beyond the user’s own entitlementsAIRM-196Insiders use AI agents to reach beyond their own access
01Governance (part)Roles & accountability · Audit & assurance
AIRM-007Segregation of duties collapses across AI config, operation, and reviewAIRM-015Inability to audit AI: no complete inventory of models, agents, and connectors
03Data & Privacy (part)Privacy & consent
AIRM-038Re-identification of individuals from aggregated or model-derived data
04Model Development (part)Accuracy & validation
AIRM-054Skipped or insufficient red-teaming, safety, and fairness testing
06Safety & Ethics (part)Harmful / inaccurate output · Misuse & dual-use
AIRM-199No guardrails on what customer-facing AI says in our nameAIRM-096Misuse of AI for disinformation, fraud, or social engineeringAIRM-097Model gives meaningful help toward weapons or cyberattack capability
07Operations (part)MLOps & deployment · Monitoring & observability · Availability & scalability · Incident management · Change & lifecycle management
AIRM-107AI tool-execution or test containment fails and AI reaches real systemsAIRM-108No monitoring, acted on, for drift, biased outcomes, or abuseAIRM-109Inadequate logging of inputs, outputs, tool calls, agent decisions, and who authorized themAIRM-110Non-determinism making anomalous behavior hard to define or detectAIRM-111Sanctioned agent activity indistinguishable from agentic attack in telemetryAIRM-114Denial-of-wallet: attackers run up AI compute costsAIRM-115Immature AI incident response and escalation processesAIRM-116Inability to investigate AI incidents due to non-determinism / poor logsAIRM-117Inadequate content-moderation / trust-&-safety operations at scaleAIRM-201No tested off-switch or fallback for a misbehaving AI featureAIRM-123Controls re-verified less often than the AI system changes
08Legal & Compliance (part)Contractual & licensing
AIRM-198Customers and partners feeding our data into their AI, with no terms or visibility
09Third-Party (part)Open-source component risk
AIRM-148Malicious or compromised open-source models / componentsAIRM-189Models deployed without verifying integrity or lineage
10Workforce (part)Societal & reputational
AIRM-160Reputational damage amplified when AI fails in public
11Financial (part)Investment / procurement
AIRM-169Insurance excludes AI losses, or defines 'user' as human only
12Autonomy (part)Autonomy, alignment & oversight · Tool access & authorization · Memory & context integrity · Multi-agent orchestration · Action integrity & traceability
AIRM-176No working interrupt or kill switch for an agent mid-taskAIRM-206No human approval before high-consequence agent actionsAIRM-177Agent granted a broader set of invokable tools / functions than its task requiresAIRM-178Agent identity misuse, impersonation, or delegated-credential abuseAIRM-179Third-party tool/connector servers as an unvetted execution surfaceAIRM-180Memory poisoning corrupting agent state across sessionsAIRM-181Context-window manipulation altering agent behaviorAIRM-184Misplaced inter-agent trust or agent collusionAIRM-186Agents able to take irreversible real-world actions
Is theirs aimed at us?
Not on the AI Risk Map, by design
Deepfake impersonation, AI-enabled fraud and phishing, autonomous attack agents, frontier capability with the guardrails stripped.
Managed on the threat register.
How the mapping changes
The mapping has its own revision number, pinned to the AI Risk Map version and the Four Faces version it was made under. A change to the mapping alone is a new revision of this page and the companion files, not a new release of the AI Risk Map. Every AI Risk Map release re-checks it.
Connects
With
Discuss
This mapping is open for comment. If a factor sits in the wrong face, say so in a GitHub issue on the AI Risk Map repository.