CompTIA SecAI+ key terms

The set of policies, roles, and oversight processes an organization uses to direct how AI systems are approved, monitored, and retired so that they stay aligned with legal, ethical, and business requirements.

A living catalog that records identified AI risks, their likelihood and impact ratings, assigned owners, and mitigation status, used to track and prioritize risk treatment across the AI portfolio.

The level of risk that remains after planned controls and mitigations have been applied, which decision makers must formally accept or reject before a system goes into production.

The amount and type of risk an organization is willing to accept in pursuit of its objectives, expressed as thresholds that guide whether an AI use case may proceed, be conditioned, or be rejected.

A short standardized document describing a model's intended use, training data, performance metrics, limitations, and ethical considerations so downstream users can judge whether it fits their context.

A structured record documenting a dataset's provenance, composition, collection method, preprocessing, and known biases, helping consumers assess suitability and legal constraints before training or evaluating models.

A structured review that evaluates how a proposed AI system may affect individuals, groups, and society, including fairness, privacy, safety, and human rights effects, before deployment is approved.

Systematic and repeatable unfairness in model outputs that disadvantages particular groups, often arising from skewed training data, flawed labeling, or proxy variables correlated with protected characteristics.

The degree to which the reasoning behind a model's output can be understood by humans, ranging from inherently interpretable models to post-hoc techniques that approximate feature contributions.

The deliberate placement of human review, approval, or intervention points in an AI workflow so that consequential decisions are not left solely to automated systems.

A control pattern in which a person reviews and can approve, modify, or reject an AI recommendation before the action is executed, preserving accountability for high-impact decisions.

The human tendency to over-trust automated output and accept it with less scrutiny than an equivalent human judgment, which can undermine the effectiveness of oversight controls.

A maintained record of every AI system, model, and third-party AI service in use, including owner, purpose, data sources, and deployment status, forming the basis for governance and monitoring.

Unauthorized or unmanaged use of AI tools and models by employees outside approved channels, creating data leakage, compliance, and visibility gaps for the organization.

An input deliberately perturbed in ways imperceptible to humans so that a model misclassifies it or produces an attacker-chosen output, exploiting the model's learned decision boundaries.

An attack performed at inference time in which crafted inputs are submitted to a deployed model to bypass detection or cause a wrong result, without altering the model itself.

Tampering with training data or the training process so that a model learns malicious behavior or hidden triggers, degrading its integrity before or during deployment.

A hidden pattern embedded during training that causes a model to produce attacker-desired output whenever the pattern appears in input, while behaving normally otherwise.

An attack in which crafted text in user input or retrieved content causes a language model to ignore its instructions and follow the attacker's, potentially triggering unintended tool actions.

A prompt injection delivered through content the model ingests, such as a web page, document, or email, rather than through the user's direct input, making it hard to spot.

A technique that circumvents a model's safety alignment or content restrictions, often through role-play or reframing, to elicit outputs the provider intended to refuse.

An attack that queries a deployed model systematically to reconstruct an approximation of its behavior or parameters, enabling theft of intellectual property or cheaper follow-on attacks.

A privacy attack that determines whether a specific record was part of a model's training data by exploiting differences in how the model responds to seen versus unseen examples.

Unauthorized transfer of sensitive data out of an environment, which in AI contexts can occur through model outputs, tool calls, or plugins manipulated by injected instructions.

Exposure created by dependencies on third-party models, datasets, libraries, and services whose provenance, integrity, or security practices are outside the organization's direct control.

Structured adversarial testing in which a team deliberately attacks a model or AI system to find safety, security, and misuse weaknesses before adversaries do.

The ability of a model to maintain correct and safe behavior when subjected to crafted inputs, distribution shift, or other adversarial conditions, often improved through adversarial training.

A runtime control that constrains model input or output, such as content filters, topic restrictions, or schema validation, to prevent harmful, off-policy, or malformed responses.

A technique that adds calibrated noise to data or query results so that outputs reveal little about any single individual, limiting the success of privacy attacks on models.

The practice of collecting and retaining only the data actually needed for a defined purpose, reducing exposure and regulatory burden when training or operating AI systems.

The documented origin and history of a dataset, including who collected it, how it was transformed, and what rights apply, used to verify trustworthiness and licensing.

Cryptographically signing model artifacts so that consumers can verify the model's origin and confirm it has not been altered since publication, protecting integrity in the supply chain.

A controlled repository that versions model artifacts and metadata, records approvals, and governs which model versions may be promoted to production environments.

Restricting how many requests a client may make in a time window, which blunts automated model extraction, credential stuffing, and resource-exhaustion attempts against inference endpoints.

Granting an AI agent or automated workflow only the permissions, data access, and tool capabilities needed for its specific task, limiting the blast radius if it is manipulated or compromised.

The capability of a model to invoke external functions, APIs, or plugins to act beyond text generation, which expands usefulness but adds attack surface for injected instructions.

Using machine learning to rank, group, and prioritize security alerts so analysts focus on the highest-risk events, with human review retained for final disposition.

Desensitization that occurs when analysts face overwhelming volumes of alerts, causing real incidents to be missed; a key problem AI-assisted triage aims to reduce.

A detection approach that baselines normal activity for users, hosts, and services and flags statistically unusual behavior that may indicate compromised accounts or insider misuse.

Identifying observations that deviate significantly from learned normal patterns, used in security monitoring for network traffic, login behavior, and process execution without relying on known signatures.

An alert or classification that incorrectly flags benign activity as malicious, consuming analyst time and eroding trust in detection systems when rates are high.

A codified sequence of automated response actions, such as enrichment, containment, and notification, that a security orchestration platform runs when defined conditions are met.

Automatically adding context such as reputation, geolocation, and known-campaign data to alerts or indicators so analysts can judge severity faster and more consistently.

Degradation in a deployed model's performance as real-world data diverges from training data over time, requiring monitoring, retraining, or replacement to keep outputs reliable.

Ongoing automated collection and review of metrics, logs, and behavior for a deployed AI system to detect drift, abuse, incidents, and control failures in near real time.

An automated pipeline that builds, tests, scans, and deploys model artifacts through controlled stages, enforcing security gates and reproducibility before production release.

The practices and tooling that operationalize machine learning delivery, covering data pipelines, training, versioning, deployment, and monitoring across the model lifecycle.

An engineering principle in which security requirements, threat modeling, and abuse-case analysis are built into an AI system from the start rather than added after deployment.