threat_id stringlengths 6 6 | threat_name stringlengths 9 42 | threat_category stringlengths 15 42 | ai_system_type stringclasses 9
values | stride_category stringlengths 9 33 | nist_ai_rmf_function stringlengths 6 32 | iso_42001_control_theme stringlengths 15 66 | mitre_atlas_technique stringlengths 9 9 | owasp_llm_reference stringclasses 10
values | description stringlengths 164 272 | attack_path stringlengths 147 263 | assessment_guidance stringlengths 136 194 | priority_generative_ai stringclasses 4
values | priority_agentic_ai stringclasses 4
values | priority_predictive_ai stringclasses 4
values | priority_supply_chain stringclasses 3
values | related_vulnerabilities stringlengths 16 34 | controls stringlengths 115 215 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
TV-001 | Prompt Injection | Input Manipulation | Generative AI; Agentic AI | Tampering; Spoofing | GOVERN; MAP; MEASURE | Threat and Vulnerability Management; Information Security Controls | AML.T0051 | LLM01 | An attacker supplies malicious instructions through user input, retrieved content, documents, webpages, messages, or tool outputs to alter model behavior. Can override intended instructions, expose sensitive information, bypass safeguards, and induce unauthorized actions. | Direct: attacker types malicious instructions in user input field. Indirect: malicious instructions embedded in documents or external content retrieved by the system. | Test direct and indirect injection paths. Assess whether untrusted content can influence decisions, outputs, or tool use. Test multimodal injection vectors. | Critical | Critical | Low | Medium | VUL-003; VUL-008; VUL-009 | Treat model instructions and untrusted content as separate trust domains; retrieval content sanitization; tool-use policies enforced outside the model; human approval for high-risk actions |
TV-002 | Data Poisoning | Training Data Attacks | Predictive AI; Generative AI | Tampering | GOVERN; MAP; MEASURE; MANAGE | Data Governance; Threat and Vulnerability Management | AML.T0020 | LLM03 | Deliberate insertion, modification, or curation of training, fine-tuning, feedback, or retrieval data to influence future model behavior. Can degrade performance broadly or create targeted backdoors. | Availability poisoning: degrade overall performance. Integrity poisoning: create targeted backdoor behavior. Bias poisoning: skew outcomes for specific groups. | Test training data validation controls. Assess poisoning in pre-training data, fine-tuning corpora, labels, retraining feedback loops, and RAG knowledge bases. | High | High | Critical | High | VUL-014; VUL-015; VUL-017; VUL-018 | Provenance verification; data quality rules; outlier detection; trusted labeling processes; holdout integrity datasets; differential retraining review |
TV-003 | Backdoor Injection | Training Data Attacks | Predictive AI; Generative AI | Tampering | GOVERN; MAP; MANAGE | Threat and Vulnerability Management; Information Security Controls | AML.T0018 | LLM03 | Hidden triggers embedded into training data or model behavior so the system acts normally most of the time but fails or behaves maliciously when the trigger appears. Difficult to detect before deployment. | Malicious training samples containing hidden trigger patterns. Model passes standard validation but contains latent malicious conditional behavior. | Evaluate outsourced training and third-party model imports. Test for suspicious trigger-response patterns. Use targeted test cases to surface hidden conditional behavior. | High | Medium | Critical | High | VUL-014; VUL-023; VUL-033 | Data provenance controls; integrity validation; third-party model verification; adversarial testing for trigger-response patterns; cryptographic signing of training artifacts |
TV-004 | Model Extraction via Queries | Intellectual Property Theft | Generative AI; Predictive AI; Agentic AI | Information Disclosure | GOVERN; PROTECT; MEASURE | Information Security Controls; Threat and Vulnerability Management | AML.T0005 | LLM10 | Attacker systematically interacts with a model API or inference service to learn its behavior and reproduce a close functional copy. Threatens intellectual property and security. | Systematic API queries using varied inputs to map decision boundaries. Statistical analysis of responses to reconstruct model behavior. Offline replication for evasion strategy development. | Examine whether repeated querying confidence outputs or detailed responses make extraction feasible. Assess rate limits and behavioral anomaly detection. | High | Medium | High | High | VUL-007; VUL-002 | Rate limiting; query monitoring; response minimization; access controls; watermarking; behavioral anomaly detection |
TV-005 | Adversarial Evasion | Model Manipulation at Inference | Predictive AI; Generative AI | Tampering | MAP; MEASURE; MANAGE | Threat and Vulnerability Management | AML.T0015 | LLM05 | Crafted inputs cause the model to misclassify, mis-rank, or generate unsafe results during inference. Relevant to fraud detection, vision, malware detection, and classification systems. | Targeted attacks: force specific misclassification. Untargeted attacks: degrade overall performance. Semantic manipulation, obfuscation, and environmental perturbation. | Test targeted and untargeted evasion, semantic manipulation, obfuscation, environmental perturbation, and sensitivity to adversarially chosen input changes. | High | Medium | Critical | Medium | VUL-003; VUL-029 | Adversarial robustness testing; input preprocessing; ensemble defenses; confidence thresholds; human review for high-risk decisions |
TV-006 | Unauthorized Tool Use | Agentic Action Abuse | Agentic AI | Tampering; Elevation of Privilege | GOVERN; PROTECT; MEASURE; MANAGE | Information Security Controls; Threat and Vulnerability Management | AML.T0051 | LLM07 | Model or agent is induced to call plugins, APIs, scripts, databases, or enterprise systems in ways that violate intended authority or business policy. Impact moves from unsafe output to unsafe action. | Prompt manipulation to trigger sensitive tool calls. Tool output manipulation to chain unauthorized actions. Multi-step planning abuse using permitted tools to achieve restricted outcomes. | Assess whether model can trigger sensitive tools through prompt manipulation, tool output manipulation, hidden argument injection, or multi-step planning abuse. | Medium | Critical | Low | High | VUL-009; VUL-010; VUL-001 | Least-privilege tool access; approval gates for sensitive actions; action sandboxing; short-lived credentials; step-level logging; budget, time, and action limits |
TV-007 | Sensitive Data Extraction | Data Exfiltration | Generative AI; Agentic AI; Predictive AI | Information Disclosure | PROTECT; MEASURE | Information Security Controls; Privacy Controls | AML.T0024 | LLM02 | Attackers recover confidential training data, personal data, secrets, business records, or proprietary knowledge from the model, its outputs, associated storage, or surrounding components. | Adversarial prompts to elicit memorized training data. API abuse to extract embedded secrets. Membership inference to confirm individual record presence. Model inversion to reconstruct sensitive attributes. | Test for leakage through ordinary interaction, API abuse, retrieval abuse, debugging interfaces, prompt replay, and model-assisted reconstruction. | Critical | Critical | High | High | VUL-004; VUL-005; VUL-026; VUL-027 | Output filtering and redaction; differential privacy; access segmentation; encryption in transit and at rest; data loss prevention on outputs; adversarial leakage testing |
TV-008 | RAG Corpus Poisoning | Retrieval Attacks | Generative AI; Agentic AI | Tampering | MAP; MEASURE; MANAGE | Data Governance; Threat and Vulnerability Management | AML.T0020 | LLM10 | Malicious or misleading content inserted into document repositories, vector databases, or enterprise knowledge sources that the model later retrieves and treats as authoritative. | Insert poisoned documents into retrieval corpus. Embed hidden instructions in retrieved content. Plant false facts or operational redirects that the model incorporates into outputs. | Test whether poisoned documents can alter output behavior, suppress correct information, induce prompt injection, or cause confidential data disclosure. | Critical | Critical | Low | High | VUL-008; VUL-014; VUL-020 | Content sanitization at retrieval boundary; source trust classification; retrieval integrity validation; test retrieval corpus for embedded instructions; human review of high-stakes retrieved outputs |
TV-009 | API Abuse | Interface Exploitation | Generative AI; Agentic AI; Predictive AI | Tampering; Denial of Service | PROTECT; MEASURE | Information Security Controls | AML.T0016 | LLM10 | Attackers exploit exposed AI interfaces to manipulate model behavior, extract data, steal models, or degrade service. High-frequency vector with direct access to the model and orchestration environment. | Weak authentication exploitation. Broken authorization bypass. Query automation for extraction. Endpoint discovery for reconnaissance. Replay abuse using captured requests. | Assess for weak authentication, broken authorization, missing rate limits, query automation, endpoint discovery, replay abuse, and insecure parameter handling. | High | Critical | High | Critical | VUL-002; VUL-007; VUL-001 | Strong authentication and authorization; rate limits and quotas; anomaly detection; endpoint protection; input validation; secure parameter handling |
TV-010 | API Token Compromise | Credential Theft | Generative AI; Agentic AI; Predictive AI | Spoofing; Information Disclosure | PROTECT | Information Security Controls | AML.T0012 | LLM09 | Attackers steal, leak, reuse, or misuse credentials granting access to AI models, tools, data stores, orchestration services, or cloud resources. Many AI environments rely on overprivileged or poorly rotated tokens. | Secret exposure in prompts, logs, code repositories, CI/CD pipelines, browser storage, and third-party integrations. Token reuse after rotation failure. | Assess secret exposure in prompts, logs, repositories, CI/CD pipelines, browser storage, and third-party integrations. Test rotation discipline. | High | Critical | High | Critical | VUL-001; VUL-002 | Secret scanning; short-lived credentials; token rotation; vault-based secret management; least privilege; audit logging of token use |
TV-011 | Third-Party Component Compromise | Supply Chain Attacks | Generative AI; Agentic AI; Predictive AI | Tampering | GOVERN; MAP | Third-Party and Supply Chain Risk | AML.T0010 | LLM05 | Attackers exploit or subvert external models, libraries, prompt frameworks, package dependencies, APIs, or model-serving components. Most organizations assemble AI systems from open-source and vendor-supplied parts. | Malicious package injection in dependencies. Backdoored pretrained model distribution. Compromised prompt framework update. Subverted vendor API with altered behavior. | Assess whether imported models, packages, and services can introduce malware, hidden behaviors, unsafe defaults, poisoned dependencies, or undisclosed data flows. | High | High | High | Critical | VUL-033; VUL-034; VUL-035 | Software bill of materials (SBOM) for AI; dependency scanning; cryptographic artifact verification; model provenance checks; vendor security assessment |
TV-012 | Parameter Tampering | Model Integrity Attacks | Generative AI; Agentic AI; Predictive AI | Tampering | PROTECT; MEASURE | Information Security Controls; Change Management | AML.T0031 | LLM05 | Attacker or unauthorized insider modifies model weights, prompts, hyperparameters, temperature settings, routing logic, safety thresholds, or decision parameters to alter system behavior. | Unauthorized modification of prompt templates. Direct model weight manipulation. Safety threshold reduction. Routing logic alteration to bypass controls. | Examine access paths to model configuration, parameter update workflows, approval controls, and whether small changes produce disproportionate security impact. | High | High | High | High | VUL-004; VUL-046; VUL-005 | Strict change control with separation of duties; immutable parameter records; cryptographic artifact signing; access restriction to configuration assets; audit logging of all parameter changes |
TV-013 | Insider Sabotage | Insider Threats | Generative AI; Agentic AI; Predictive AI | Tampering; Repudiation | GOVERN; PROTECT | Information Security Controls; Access Control | AML.T0031 | LLM06 | Authorized personnel intentionally degrade, corrupt, or weaponize the AI system by introducing dormant logic, malicious code, bad data, or harmful operational changes. Developers and MLOps personnel have broad access. | Implanted dormant failure logic. Malicious training data injection by insider. Prompt template modification. Monitoring suppression or alert disabling. | Assess whether insider actions could implant delayed failures, poison training data, change prompts, weaken monitoring, or suppress alerts without timely detection. | High | High | High | High | VUL-001; VUL-004; VUL-030 | Separation of duties; least privilege; behavioral monitoring; peer review for sensitive changes; immutable audit logs; background screening and access reviews |
TV-014 | Insider Subversion | Insider Threats | Generative AI; Agentic AI; Predictive AI | Spoofing; Information Disclosure | GOVERN; PROTECT | Information Security Controls; Access Control | AML.T0031 | LLM06 | Internal personnel are bribed, coerced, or recruited to steal AI assets, leak data, or manipulate system behavior for external actors. Objective includes espionage or competitive theft. | Exfiltration of model weights or proprietary training data. Leak of system prompts or evaluation sets. Strategic compromise of AI governance artifacts for competitive intelligence. | Examine privileged access, separation of duties, behavioral anomalies, unusual artifact access, and whether sensitive model assets can be exported by a small number of insiders. | High | High | High | Critical | VUL-001; VUL-006; VUL-040 | Data loss prevention; exfiltration monitoring; dual control for high-value artifact access; behavioral anomaly detection; clear asset classification and handling policies |
TV-015 | Data Provenance Falsification | Data Integrity Attacks | Generative AI; Predictive AI | Tampering; Repudiation | MAP; MEASURE | Data Governance | AML.T0020 | LLM03 | Metadata, lineage records, ownership fields, timestamps, source identifiers, or chain-of-custody records are altered to disguise the true origin or integrity of AI data. | Overwrite source metadata to legitimize poisoned dataset. Forge timestamp records to conceal late data modification. Detach lineage records from actual training artifacts. | Assess whether source records can be forged, overwritten, or detached from actual datasets and whether data trust decisions rely too heavily on editable metadata. | Medium | Medium | High | High | VUL-012; VUL-018 | Cryptographic signing of datasets and lineage records; immutable provenance logs; independent provenance validation; separation of metadata write access |
TV-016 | Label Poisoning | Training Data Attacks | Predictive AI | Tampering | MAP; MEASURE; MANAGE | Data Governance; Threat and Vulnerability Management | AML.T0020 | LLM03 | Labels in supervised learning datasets are manipulated, corrupted, or systematically skewed to alter model decision boundaries and degrade reliability or create targeted blind spots. | Systematic label corruption for targeted class. Random label noise to degrade overall accuracy. Reviewer manipulation to introduce systematic annotation bias. | Assess annotation workflows, reviewer independence, class distribution anomalies, suspicious relabeling events, and whether label quality is monitored throughout retraining. | Low | Low | Critical | High | VUL-014; VUL-017; VUL-015 | Trusted labeling processes; multi-reviewer consensus; label distribution monitoring; holdout integrity datasets; anomaly detection on label changes |
TV-017 | Bias Exploitation Through Imbalanced Data | Fairness Attacks | Predictive AI; Generative AI | Tampering | MAP; MEASURE | Fairness and Bias Controls; Data Governance | AML.T0020 | LLM09 | Attackers or negligent processes exploit underrepresented groups, skewed classes, or socially biased data distributions to produce discriminatory or harmful outcomes in high-stakes decisions. | Intentionally amplify biased or nonrepresentative inputs to steer outcomes. Exploit known subgroup underrepresentation in training data to produce systematically wrong results for target groups. | Assess representativeness, subgroup error rates, data collection bias, and whether adversaries could steer outcomes by amplifying biased inputs. | High | Medium | Critical | High | VUL-015; VUL-017; VUL-023 | Bias testing across subgroups; representative dataset requirements; fairness metrics in evaluation; subgroup performance monitoring; bias audit before deployment |
TV-018 | Model Inversion | Privacy Attacks | Predictive AI; Generative AI | Information Disclosure | PROTECT; MEASURE | Privacy Controls; Information Security Controls | AML.T0024 | LLM02 | Attacker analyzes model responses to reconstruct sensitive attributes, representative records, or approximations of training data. Relevant where models are trained on healthcare, biometric, or financial data. | Systematic API queries analyzing confidence outputs to reconstruct sensitive records. Gradient analysis to infer private attributes. Embedding analysis to reverse-engineer training examples. | Assess whether outputs, gradients, embedding access, or repeated targeted queries enable inference of private records or sensitive attributes. | High | Medium | High | Medium | VUL-007; VUL-027 | Output minimization; differential privacy; confidence value restrictions; access controls on embeddings and gradients; adversarial testing for reconstruction |
TV-019 | Membership Inference | Privacy Attacks | Predictive AI; Generative AI | Information Disclosure | PROTECT; MEASURE | Privacy Controls | AML.T0024 | LLM02 | Attacker determines whether a specific individual, record, or item was included in a model's training data. Creates serious privacy and legal exposure when participation in a dataset is itself sensitive. | Confidence score analysis comparing outputs for records known to be in training data versus not. Differential behavior probing for memorized examples. Shadow model training for statistical inference. | Examine whether output confidence, overfitting, differential behavior, or verbose responses allow adversaries to infer dataset membership with meaningful accuracy. | High | Medium | High | Medium | VUL-007; VUL-026; VUL-027 | Differential privacy in training; output confidence restrictions; model evaluation for memorization; access controls; regular privacy audits |
TV-020 | Gradient Leakage | Privacy Attacks in Federated Learning | Predictive AI | Information Disclosure | PROTECT; MEASURE | Privacy Controls; Information Security Controls | AML.T0024 | LLM09 | Attacker reconstructs training examples or infers sensitive information from gradient updates or model parameter changes shared during distributed or federated learning. | Gradient inversion attacks reconstructing input data from shared updates. Statistical analysis of gradient distributions to infer data characteristics. Targeted inference using gradient sensitivity. | Evaluate secure aggregation, differential privacy, clipping, update access, and whether shared training signals could reveal individual data points. | Low | Low | High | High | VUL-026; VUL-027 | Secure aggregation; differential privacy with gradient clipping; restricted gradient sharing; anomaly detection on gradient distributions |
TV-021 | Hallucination Exploitation | Output Manipulation | Generative AI | Tampering; Information Disclosure | MEASURE; MANAGE | Threat and Vulnerability Management; Information Security Controls | AML.T0054 | LLM09 | Attackers intentionally cause a generative model to produce false, fabricated, or misleading content used to deceive users, justify action, or contaminate downstream workflows. | Craft prompts that elicit confident but false factual claims. Induce fabricated citations, credentials, or policy interpretations. Target automated systems that accept AI output without verification. | Assess whether model can be induced to invent facts, credentials, citations, procedures, or policy interpretations that materially affect operations or decisions. | Critical | High | Low | Medium | VUL-030; VUL-003 | Grounding mechanisms; verification checks; confidence indicators; output validation; restrictions on automated use of unverifiable outputs; human review gates |
TV-022 | Toxicity Induction | Output Safety Attacks | Generative AI | Tampering | MEASURE; MANAGE | Threat and Vulnerability Management | AML.T0051 | LLM01 | Attackers provoke model into generating hateful, abusive, sexually explicit, extremist, or otherwise harmful content. Creates immediate legal, reputational, and trust consequences even without system compromise. | Direct prompts requesting harmful content. Role-play and persona manipulation to bypass safety filters. Paraphrase, coded language, and multilingual obfuscation to evade detection. | Test whether adversaries can elicit toxic output across languages, contexts, and obfuscation methods including role-play, paraphrase, and coded language. | Critical | High | Low | Medium | VUL-003; VUL-028 | Output safety classifiers; content moderation layers; red team testing across languages and obfuscation methods; usage monitoring and abuse reporting |
TV-023 | Dual-Use or Malicious Repurposing | Intentional Misuse | Generative AI; Agentic AI | Tampering; Elevation of Privilege | GOVERN; MANAGE | Intended Use Controls; Threat and Vulnerability Management | AML.T0054 | LLM06 | Model designed for benign enterprise use is repurposed, stolen, or adapted for fraud, misinformation, surveillance, phishing, deepfakes, or other harmful purposes by insiders or external actors. | Employee repurposes enterprise copilot to generate phishing content. External actor obtains API access to generate synthetic fraud documents. Fine-tuned model variant used for harmful content at scale. | Assess abuse patterns, policy restrictions, customer and employee monitoring, and whether model capabilities create foreseeable misuse channels. | High | High | Medium | High | VUL-031; VUL-028; VUL-001 | Use policy enforcement; behavioral monitoring; capability restrictions; access controls; abuse detection and reporting channels |
TV-024 | Overreliance / Automation Bias | Human Factor Threats | Generative AI; Agentic AI; Predictive AI | Repudiation | GOVERN; MANAGE | Human Oversight Controls | AML.T0054 | LLM09 | Humans accept AI outputs or recommendations with insufficient scrutiny, leading to poor decisions, unsafe approvals, or unchecked propagation of model error. Even a technically accurate system can cause harm if unchecked. | Users defer to model output in high-stakes decisions without independent verification. Automated downstream processes treat AI output as authoritative. Escalation paths bypassed due to assumed model accuracy. | Assess whether users are likely to defer to the model in high-stakes decisions and whether process controls force independent verification where needed. | Critical | Critical | High | Medium | VUL-030; VUL-010; VUL-032 | Human-in-the-loop controls; escalation thresholds; decision audit requirements; user training on AI limitations; confidence and uncertainty communication in outputs |
TV-025 | Shadow AI Use | Unauthorized AI Deployment | Generative AI; Predictive AI | Spoofing; Information Disclosure | GOVERN | Intended Use Controls; Information Security Controls | AML.T0012 | LLM06 | Employees or business units introduce unapproved AI tools, models, or services outside security, compliance, and architecture review. Exposes organizations to uncontrolled data transfer and unmonitored decision-making. | Employee uses external copilot or browser plugin with sensitive business data. Business unit deploys unapproved SaaS AI without security review. Local model used on unmanaged device for regulated workflows. | Determine whether staff are using external copilots, browser plugins, SaaS models, or local agents without authorization, and whether sensitive data is being routed to unsanctioned systems. | High | High | Medium | High | VUL-031; VUL-001; VUL-034 | AI use policy; asset discovery; network monitoring for unauthorized AI endpoints; browser extension controls; employee awareness and reporting channels |
TV-026 | Excessive Agency Abuse | Agentic Action Abuse | Agentic AI | Elevation of Privilege; Tampering | GOVERN; PROTECT; MANAGE | Information Security Controls; Human Oversight Controls | AML.T0051 | LLM08 | Model or agent with excessive permissions or autonomy is induced to perform actions beyond intended scope. Combination of autonomous planning, tool access, and permissive integration creates high-impact action paths. | Prompt manipulation induces agent to write, delete, or transact outside intended scope. Multi-step planning chains permitted actions to achieve restricted outcomes. Permissive tool integration exploited to escalate privileges. | Assess whether agent can write, delete, transact, message, escalate, or reconfigure systems without independent authorization or human review. | Low | Critical | Low | High | VUL-009; VUL-010; VUL-006 | Least-privilege tool access; approval gates; action sandboxing; short-lived credentials; step-level logging; budget, time, and action limits; scope restriction policies |
TV-027 | Agent Collusion | Multi-Agent Attacks | Agentic AI | Tampering; Elevation of Privilege | GOVERN; PROTECT; MEASURE | Information Security Controls; Threat and Vulnerability Management | AML.T0054 | LLM08 | Multiple AI agents coordinate, intentionally or emergently, to manipulate decisions, bypass controls, or amplify harmful outcomes. One compromised agent can influence others through hidden coordination pathways. | Compromised orchestrator agent directs subagents to bypass controls. Emergent coordination between agents creates unintended harmful feedback loops. Agent communication channels exploited for lateral movement. | Assess whether a compromised or malicious agent can influence other agents, create harmful feedback loops, or distribute unsafe actions across multiple actors to evade detection. | Low | Critical | Low | High | VUL-009; VUL-041; VUL-001 | Strong agent identity and authentication; communication integrity controls; role separation between agents; behavioral monitoring for multi-agent coordination; policy enforcement at orchestration layer |
TV-028 | Denial of Service / Denial of Wallet | Resource Exhaustion Attacks | Generative AI; Agentic AI; Predictive AI | Denial of Service | PROTECT; MANAGE | Information Security Controls; Resilience Controls | AML.T0029 | LLM04 | Attackers exhaust compute, token, memory, concurrency, storage, or budget resources of an AI system, reducing availability or sharply increasing cost. Generative and agentic systems enable low-volume high-cost attacks. | Crafted prompts maximizing token generation. Long tool chains triggered to exhaust orchestration resources. Worst-case inference behavior forced to increase cost. Concurrent session flooding for agentic services. | Evaluate flood resistance, concurrency control, token budgets, loop limits, spend alerts, and graceful degradation under abusive demand. | High | Critical | Medium | Medium | VUL-007; VUL-044; VUL-045 | Rate limits and per-user quotas; token budgets; circuit breakers; cost-aware request validation; concurrency limits; spend monitoring and alerts; graceful degradation modes |
TV-029 | Eavesdropping on Inputs | Interception Attacks | Generative AI; Agentic AI; Predictive AI | Information Disclosure | PROTECT | Information Security Controls; Privacy Controls | AML.T0012 | LLM02 | Attackers intercept user prompts, uploaded files, sensor streams, or transaction data before it is processed by the AI system, exposing sensitive business or personal information. | Network interception of unencrypted prompt traffic. Endpoint compromise capturing keystrokes or clipboard content before transmission. Proxy-based interception of file uploads. | Assess network encryption, endpoint compromise exposure, browser and proxy exposure, and whether model input channels are protected in transit and at collection points. | High | High | High | Medium | VUL-004; VUL-002 | TLS encryption for all input channels; endpoint security; secure input collection practices; network segmentation; monitoring for unauthorized proxy or interception activity |
TV-030 | Eavesdropping on Outputs | Interception Attacks | Generative AI; Agentic AI; Predictive AI | Information Disclosure | PROTECT | Information Security Controls; Privacy Controls | AML.T0012 | LLM02 | Attackers intercept model responses, decision results, generated content, confidence values, or tool results as they leave the AI system, exposing confidential business logic or personal data. | Interception of unencrypted API responses. Browser rendering path compromise capturing displayed outputs. Inter-service messaging interception in microservice architectures. | Assess output channels, logging systems, browser rendering paths, inter-service messaging, and whether outputs are protected in transit and at rest. | High | High | High | Medium | VUL-004; VUL-005 | TLS encryption for all output channels; output logging access controls; secure inter-service communication; browser security controls; monitoring for unauthorized interception |
TV-031 | Espionage Against AI Assets | Competitive and State-Level Threats | Generative AI; Predictive AI | Information Disclosure | PROTECT | Information Security Controls; Access Control | AML.T0012 | LLM06 | Attackers infiltrate the organization or its suppliers to steal training data, model artifacts, fine-tuning sets, prompts, evaluation results, or strategic AI plans for competitive or strategic advantage. | Long-term insider access to model repositories. Supply chain infiltration to access proprietary training data. Covert exfiltration of fine-tuned model artifacts. Theft of evaluation results revealing capability details. | Examine insider access, exfiltration paths, artifact repositories, data lake exposure, and whether attackers could quietly study or remove high-value AI assets over time. | High | High | High | Critical | VUL-006; VUL-013; VUL-040 | Data loss prevention; exfiltration monitoring; artifact access controls; security clearance for sensitive AI assets; anomaly detection on repository access; encryption of high-value artifacts |
TV-032 | Physical Tampering | Physical Security Threats | Generative AI; Agentic AI; Predictive AI | Tampering | PROTECT | Physical Security Controls | AML.T0012 | LLM09 | Attackers manipulate hardware, storage media, networking equipment, edge devices, or hosting infrastructure to alter, disable, or exfiltrate AI system components. Higher likelihood in edge and industrial deployments. | Physical access to edge device to extract model weights. Storage media removal for offline analysis. Network equipment tampering to redirect model traffic. Local console access to modify system configuration. | Assess hardware access controls, removable media exposure, local console protection, environmental security, and whether physical interference can change model behavior or reveal sensitive data. | Low | Medium | Medium | High | VUL-048; VUL-049 | Physical access controls; device encryption; tamper-evident hardware; removable media controls; environmental monitoring; secure boot and firmware integrity verification |
TV-033 | Hardware Trojan Insertion | Hardware Supply Chain Attacks | Generative AI; Predictive AI | Tampering | GOVERN; PROTECT | Third-Party and Supply Chain Risk; Physical Security Controls | AML.T0010 | LLM09 | Malicious logic or hidden backdoors introduced into GPUs, accelerators, sensors, firmware, or other hardware components used by AI systems. Difficult to detect and can bypass many software-layer controls. | Hardware backdoor inserted during manufacturing. Malicious firmware update introducing hidden behavior. Side channel exfiltration of model weights through modified hardware. | Assess trusted hardware sourcing, firmware integrity, manufacturing provenance, hardware attestation, and anomalous low-level behavior indicating embedded compromise. | Low | Low | Low | Critical | VUL-048; VUL-049; VUL-033 | Trusted hardware sourcing with vendor assurance; firmware integrity verification; hardware attestation; anomaly monitoring at hardware layer; supply chain security requirements for hardware vendors |
TV-034 | Fault Injection | Physical and Environmental Attacks | Predictive AI; Agentic AI | Tampering | PROTECT; MANAGE | Physical Security Controls; Resilience Controls | AML.T0012 | LLM09 | Attackers induce errors through voltage changes, heat, clock manipulation, sensor interference, malformed inputs, or environmental manipulation to cause AI system malfunction. Relevant to embedded and industrial AI. | Voltage glitching to corrupt model inference. Sensor data manipulation to produce incorrect classifications. Malformed input sequences designed to trigger undefined behavior. | Test resilience to corrupted inputs, abnormal operating conditions, fail-safe behavior, and whether induced faults can cause silent misclassification rather than visible shutdown. | Low | Medium | Low | High | VUL-050; VUL-048 | Fault detection and safe degradation; environmental monitoring; input validation and sanitization; watchdog timers; fail-safe defaults on hardware fault detection |
TV-035 | Neglected Patching Exploitation | Software Vulnerability Exploitation | Generative AI; Agentic AI; Predictive AI | Tampering | PROTECT | Information Security Controls | AML.T0012 | LLM09 | Attackers take advantage of unpatched frameworks, runtimes, libraries, model-serving components, notebooks, or operating systems supporting AI workflows. AI ecosystems depend on fast-moving open-source packages. | Known CVE exploitation in unpatched ML serving framework. Vulnerable Jupyter notebook environment exploitation. Outdated container runtime exploitation in model serving infrastructure. | Assess patch latency, unsupported components, exposed CVEs in ML tooling, upgrade discipline, and whether security updates are blocked by fragile model pipelines. | High | High | High | High | VUL-035; VUL-011 | Patch management process covering AI-specific tooling; dependency scanning; container image scanning; upgrade testing in staging; SBOM maintenance |
TV-036 | Functional Extraction | Intellectual Property Theft | Generative AI; Predictive AI | Information Disclosure | PROTECT; MEASURE | Information Security Controls | AML.T0005 | LLM10 | Attackers create an offline model that behaves similarly enough to the target system to support attack development, policy evasion, or competitive substitution. Emphasizes reproducing operational behavior rather than exact weights. | Systematic API interaction to map input-output behavior for a target task. Statistical modeling of response patterns to approximate decision logic. Cloned model used offline to develop evasion strategies. | Assess whether system reveals enough output structure, determinism, and behavioral consistency for attackers to clone its utility for downstream offensive use. | High | Medium | High | High | VUL-007; VUL-002 | Query monitoring; rate limiting; response minimization; behavioral anomaly detection; model watermarking; output confidence restrictions |
TV-037 | Black-Box Manipulation | Model Exploitation Without Internal Access | Generative AI; Predictive AI | Tampering | MAP; MEASURE | Threat and Vulnerability Management | AML.T0016 | LLM05 | Attackers exploit opacity of a model to probe its behavior, infer weaknesses, and craft attacks without needing internal access to its architecture or weights. Lack of interpretability aids attackers. | Trial-and-error interaction to identify decision boundary blind spots. Systematic probing to find policy inconsistencies. Unstable region identification for targeted evasion or injection. | Test whether an attacker can systematically identify blind spots, unstable regions, or policy inconsistencies through interaction alone. | High | Medium | High | Medium | VUL-037; VUL-007 | Input monitoring for probing patterns; rate limiting; anomaly detection on systematic query behavior; behavioral consistency testing; explainability controls for internal audit |
TV-038 | Model Drift Exploitation | Temporal Attacks | Predictive AI; Generative AI | Tampering | MEASURE; MANAGE | Threat and Vulnerability Management; Monitoring Controls | AML.T0016 | LLM09 | Attackers take advantage of model misalignment with current data, behavior, or environmental conditions. Adversaries can intentionally steer or time attacks to exploit periods when model is least calibrated. | Timing attacks exploiting known retraining cycles when model is most stale. Deliberate steering of data distribution to accelerate drift. Exploiting known performance degradation on recent data patterns. | Determine whether organization can detect drift quickly, isolate its effects, and prevent attackers from exploiting known stale behavior in production. | Low | Low | Critical | Medium | VUL-039; VUL-023 | Continuous drift monitoring; automated drift alerts; short retraining cycles for high-risk models; performance monitoring with threshold-based review triggers |
TV-039 | Generalization Failure Exploitation | Out-of-Distribution Attacks | Predictive AI; Generative AI | Tampering | MAP; MEASURE | Threat and Vulnerability Management | AML.T0016 | LLM05 | Attackers capitalize on overfitting, underfitting, brittle boundaries, or narrow training coverage to force wrong model behavior on novel but realistic inputs. Deliberately targets out-of-distribution data. | Systematic search for inputs outside training distribution that produce high-confidence wrong outputs. Edge case construction targeting underrepresented scenarios. Real-world data drift into territory model handles poorly. | Assess edge-case exploration, subgroup testing, out-of-domain inputs, and whether attackers can reliably trigger failure on data outside standard evaluation sets. | High | Medium | Critical | Medium | VUL-023; VUL-024; VUL-039 | Adversarial out-of-distribution testing; uncertainty quantification; confidence calibration; subgroup evaluation; robust training data coverage requirements |
TV-040 | Transparency Deficit Exploitation | Governance Attacks | Generative AI; Predictive AI; Agentic AI | Repudiation | GOVERN; MEASURE | Explainability and Transparency Controls | AML.T0054 | LLM09 | Attackers or negligent actors benefit from organization's inability to explain, justify, or audit model decisions. Hides biased outcomes, obscures manipulated behavior, and delays incident response. | Exploit lack of explainability to conceal discriminatory outcomes. Manipulate model behavior knowing audit capability is insufficient to detect the change. Delay incident response by ensuring outputs cannot be traced. | Assess whether lack of explainability creates operational blind spots that attackers can exploit or that prevent teams from understanding when the AI system has been manipulated. | High | High | High | Medium | VUL-037; VUL-025; VUL-005 | Explainability controls proportional to use case risk; output traceability; audit logging; regular explainability audits; board-level reporting on model accountability |
TV-041 | Homogenization Risk Exploitation | Systemic Risk Attacks | Generative AI; Predictive AI | Tampering | GOVERN; MAP | Third-Party and Supply Chain Risk | AML.T0010 | LLM09 | Attackers target widely adopted models, dependencies, or architectural patterns knowing a single exploit path may affect many systems at once. Creates systemic risk through AI monocultures. | Single vulnerability in widely used foundation model exploited across multiple dependent enterprise deployments. Common prompt framework compromise affecting all users simultaneously. Shared dependency poisoning affecting many downstream systems. | Review dependence on common models, shared third-party services, uniform prompt frameworks, and whether a single compromise could propagate broadly through the environment. | High | High | High | Critical | VUL-033; VUL-034; VUL-041 | Model diversity strategy; concentration risk assessment; alternative sourcing for critical AI capabilities; monitoring for cross-system impact from common dependency vulnerabilities |
TV-042 | Indirect Prompt Injection | Input Manipulation | Generative AI; Agentic AI | Tampering | PROTECT; MEASURE | Threat and Vulnerability Management; Information Security Controls | AML.T0051 | LLM01 | Malicious instructions embedded in external content that the model later reads as part of retrieval, browsing, search, email processing, document parsing, or task execution. No direct user session access required. | Malicious instructions embedded in a retrieved document that override system prompt. Hidden commands in email body that redirect agent behavior. Hostile content in a ticket or wiki article that triggers data exfiltration. | Test whether hostile content in documents, tickets, code comments, wikis, or websites can alter behavior, exfiltrate data, or trigger unauthorized actions. | Critical | Critical | Low | High | VUL-008; VUL-003; VUL-009 | Content sanitization at retrieval boundary; trust domain separation between instructions and data; output validation; tool-use policies enforced outside the model; human review for high-risk retrieval-based actions |
TV-043 | Tool Output Manipulation | Agentic Action Abuse | Agentic AI | Tampering | PROTECT; MEASURE | Information Security Controls; Threat and Vulnerability Management | AML.T0051 | LLM08 | Attackers poison, spoof, or compromise outputs returned from APIs, web retrieval, databases, or enterprise tools that an AI system relies on. Malicious tool output misleads planning and triggers dangerous calls. | Spoof API response to redirect agent action. Poison database query result to alter planning logic. Forge tool output to create false operational picture trusted by the model. | Assess whether system authenticates tool responses, validates schemas, scores source trust, and separates data returned by tools from instructions. | Low | Critical | Low | High | VUL-009; VUL-003; VUL-008 | Tool response authentication; schema validation for tool outputs; source trust scoring; separation of tool data and instruction processing; monitoring for anomalous tool call patterns |
TV-044 | Memory Poisoning | Agentic Persistence Attacks | Agentic AI | Tampering | PROTECT; MEASURE | Information Security Controls; Threat and Vulnerability Management | AML.T0051 | LLM08 | Attackers insert malicious instructions, false facts, hidden goals, or misleading context into an agent's persistent or semi-persistent memory. Compromise persists across sessions and influences future actions. | Inject malicious instructions into agent long-term memory store. Plant false operational facts that persist across sessions. Override prior memory entries with manipulated context. | Examine what can be written to memory, how memory is reviewed, how long it persists, and whether durable memory can override policy or trusted context. | Low | Critical | Low | High | VUL-008; VUL-009; VUL-001 | Memory write access controls; memory content validation; memory TTL and rotation policies; human review for high-impact memory updates; audit logging of memory reads and writes |
TV-045 | Goal Hijacking | Agentic Objective Manipulation | Agentic AI | Tampering; Elevation of Privilege | GOVERN; PROTECT; MANAGE | Information Security Controls; Human Oversight Controls | AML.T0051 | LLM08 | Attacker causes an agent to reinterpret its objective, optimize for attacker-favored outcomes, or deprioritize safety and policy constraints through prompt manipulation, malicious context, or task reframing. | Reframe agent task through adversarial context to shift optimization target. Use retrieved content to introduce competing objective that overrides original goal. Exploit long reasoning chains to accumulate goal drift. | Test whether system can be induced to redefine success, pursue side effects, or treat restricted actions as instrumental to accomplishing a broader task. | Low | Critical | Low | High | VUL-010; VUL-009; VUL-030 | Goal specification validation; human approval for objective-level changes; monitoring for goal drift in long-horizon tasks; constraint enforcement outside model reasoning; approval gates for high-impact action paths |
TV-046 | Autonomous Action Chaining Abuse | Agentic Action Abuse | Agentic AI | Tampering; Elevation of Privilege | GOVERN; PROTECT; MANAGE | Information Security Controls; Human Oversight Controls | AML.T0051 | LLM08 | Attackers exploit system's ability to plan and execute sequences of steps that are individually permitted but collectively harmful. Multi-step actions cross trust boundaries or combine benign tools into harmful outcomes. | Chain individually permitted tool calls to achieve restricted aggregate outcome. Exploit multi-step planning to cross authorization boundary without any single unauthorized action. Use agent's planning horizon to accumulate harm across many small permitted steps. | Evaluate whether system reasons over cumulative impact, enforces business constraints across steps, and detects suspicious action sequences. | Low | Critical | Low | High | VUL-009; VUL-010; VUL-006 | Cumulative action impact monitoring; cross-step constraint enforcement; action sequence auditing; hard limits on action chain length; human review triggers for multi-step plans above risk threshold |
TV-047 | Context Window Flooding | Input Manipulation | Generative AI; Agentic AI | Denial of Service; Tampering | PROTECT; MEASURE | Information Security Controls; Resilience Controls | AML.T0029 | LLM04 | Attackers overload model's context with large, distracting, conflicting, or adversarially ordered content to suppress trusted instructions or increase confusion. Improves success rate of injection or evasion attacks. | Flood context with irrelevant content to push system prompt toward truncation boundary. Fill context with conflicting instructions to create ambiguity. Use adversarially ordered content to suppress safety instructions. | Examine context prioritization, truncation rules, token budgeting, and whether trusted instructions remain dominant under adversarially large input loads. | High | High | Low | Medium | VUL-003; VUL-008; VUL-044 | Context prioritization rules; trusted instruction anchoring; token budget enforcement; input size limits; monitoring for abnormally large context submissions |
TV-048 | Unsafe Content Repurposing | Output Safety Attacks | Generative AI; Agentic AI | Tampering | GOVERN; MANAGE | Intended Use Controls; Threat and Vulnerability Management | AML.T0054 | LLM06 | Model used to generate phishing messages, malware-adjacent scripts, disinformation, fraudulent documents, social engineering content, or deepfake support materials. | Enterprise AI used by employee to generate phishing email templates. Model steered to produce realistic fraudulent financial documents. AI-generated social engineering scripts created for external fraud campaigns. | Assess whether misuse patterns can be detected, use restrictions are enforced, and whether model can be steered into harmful assistance despite policy controls. | High | High | Medium | High | VUL-031; VUL-028; VUL-001 | Output classification for harmful use patterns; use policy enforcement; behavioral monitoring; abuse detection and reporting channels; capability restrictions for high-risk output types |
TV-049 | Synthetic Identity and Deepfake Enablement | Fraud and Disinformation | Generative AI | Tampering; Information Disclosure | GOVERN; MANAGE | Intended Use Controls; Threat and Vulnerability Management | AML.T0054 | LLM06 | AI systems used to create realistic fake personas, voice clones, forged images, or impersonation content that supports fraud or disinformation. Materially increases social engineering effectiveness. | Generate realistic synthetic identity documents for fraud. Create voice clones for executive impersonation attacks. Produce deepfake video or images for disinformation campaigns. | Consider how easily model can generate impersonation content, what safeguards exist, and how organization monitors for abuse of these capabilities. | High | High | Low | High | VUL-031; VUL-028 | Capability restrictions on synthesis features; output watermarking; abuse monitoring; use policy enforcement; regulatory compliance controls for identity and biometric data handling |
TV-050 | Weak Accountability Exploitation | Governance Attacks | Generative AI; Agentic AI; Predictive AI | Repudiation | GOVERN | Human Oversight Controls; Incident Response | AML.T0054 | LLM09 | Attackers or negligent parties exploit unclear role assignments, missing audit trails, or weak accountability structures to conduct harmful actions without detection, attribution, or consequence. | Manipulate system exploiting absence of defined owner for AI security controls. Conduct data exfiltration knowing audit trail gaps prevent attribution. Exploit unclear escalation paths to delay incident response. | Assess whether accountability gaps allow harmful actions to persist without detection and whether responsibility assignment is clear and enforced across the AI lifecycle. | High | High | High | High | VUL-032; VUL-033; VUL-005 | Clear role and accountability assignment; complete audit trails; incident response with defined ownership; regular accountability reviews; board-level AI governance reporting |
AI Threat Taxonomy and Vulnerability Registry
Dataset Description
This dataset provides a practitioner-grade, machine-readable taxonomy of AI threat vectors and AI vulnerabilities for use in red teaming, risk assessment, model governance, and regulatory compliance programs.
The taxonomy distinguishes precisely between:
- Threat vectors: paths or mechanisms an attacker, insider, or negligent actor uses to exploit an AI environment
- Vulnerabilities: weaknesses in design, control, architecture, process, or implementation that threat vectors exploit
Coverage
- 50+ AI threat vectors organized by attack type and AI system category
- 55+ AI vulnerabilities organized by control domain
- STRIDE-AI mapping across all six threat categories
- Priority matrices for generative AI, agentic AI, predictive AI, and supply chain threats
- Assessment guidance for practitioners conducting red team and risk reviews
- Framework cross-references: ISO/IEC 42001, NIST AI RMF, MITRE ATLAS, OWASP LLM Top 10, EU AI Act
Intended Uses
- AI red team scoping and threat modeling
- AI risk register construction
- Regulatory gap assessment (EU AI Act, ISO 42001, NIST AI RMF)
- Model governance and third-party AI vendor assessment
- Executive education in AI governance and compliance
Author
Prof. Hernan Huwyler, MBA CPA CIAO
Academic Director, IE Law School Center for Risk and Compliance
Senior AI Risk and Governance Manager, Capgemini Applied AI Lab
Two-time recipient, IE Lifelong Learning Academic Excellence Award
(Management Program, 2023 and 2024)
License
Creative Commons Attribution 4.0 International (CC BY 4.0)
- Downloads last month
- 47