CompTIA SecAI+ flashcards

33 cards. Read the term and the definition here. The deck at the top of this page is optional.

NIST AI RMF functions

Four functions: GOVERN (continuous, cross-cutting), MAP (context and risk identification), MEASURE (analysis and testing), MANAGE (treatment and monitoring). They form an iterative lifecycle.

EU AI Act high-risk duties

High-risk systems require conformity assessment, technical documentation, human oversight, logging, registration in an EU database, and post-market monitoring before market placement.

ISO/IEC 42001

Certifiable standard for an AI management system, modeled on ISO 27001, covering AI policy, roles, risk treatment, internal audit, and continual improvement.

DPIA trigger

A Data Protection Impact Assessment is required when processing likely poses high risk to individuals, such as AI profiling or large-scale inference on sensitive attributes.

Model card contents

A document stating a model's intended use, training data sources, evaluation metrics and results, limitations, and ethical considerations for a specific model version.

Accountability in AI governance

Assigning a named risk owner per model and documenting approval gates ensures auditability; unowned models are a frequent audit finding.

Indirect prompt injection

Malicious instructions embedded in retrieved documents, web pages, or tool outputs that the model treats as trusted context and may follow.

Jailbreak versus prompt injection

Jailbreaks bypass safety alignment to elicit prohibited content; prompt injection subverts the instruction hierarchy to override developer intent. Mitigations differ.

Backdoor data poisoning

Planting trigger-activated behavior in training data so the model misbehaves only on specific inputs, evading clean-validation testing.

Model extraction attack

Querying a model API extensively to train a surrogate clone, stealing functionality; rate limiting and output watermarking raise attacker cost.

Membership inference attack

Determining whether a specific record was part of a model's training data, which can leak sensitive information about individuals.

Adversarial example transferability

Perturbations crafted on a surrogate model often misclassify a different black-box target, enabling attacks without direct model access.

Excessive agency risk

An LLM agent with broad write or execute permissions converts a prompt injection into real harmful actions; scope tool credentials to least privilege.

Model weight supply chain risk

Loading untrusted model files can execute code via pickle deserialization; verify provenance, signatures, and use safe serialization formats.

RAG retrieval access control

Authorization must be enforced per document at retrieval time; a shared vector index without filtering leaks documents across tenants.

Differential privacy epsilon

A privacy-loss parameter: smaller epsilon gives stronger privacy guarantees and typically lower model utility; larger epsilon does the reverse.

Confidential computing for inference

Trusted execution environments isolate prompts and model weights during processing so the host operator cannot read sensitive data in use.

PII redaction at egress

Removing or masking personal data before prompts leave the trust boundary, enforced by data loss prevention on outbound API calls.

Fail-closed guardrails

When a safety check errors or is uncertain, high-impact actions are blocked; fail-open is acceptable only for low-risk read paths.

Vector store poisoning

Injecting malicious or false content into a retrieval index so the model surfaces attacker-controlled text; mitigate by validating and tagging chunks.

Instruction and data separation

Structuring prompts so untrusted content is clearly delimited from system instructions, reducing the model's tendency to follow injected commands.

Hallucinated IOC verification

Any indicator, hash, or CVE an AI assistant produces must be checked against authoritative threat intelligence before defensive action.

Human approval for response

Destructive automated actions such as account disable or host isolation require human sign-off or a deterministic policy engine, not model discretion.

UEBA baseline drift

User behavior baselines shift as environments change; scheduled retraining and drift alerts prevent spikes in false positives.

Adversarial evasion of detectors

Attackers mutate artifacts until ML classifiers score them benign; ensembles plus rule-based backstops reduce successful evasion in security operations.

AI triage audit logging

Recording prompt, retrieved context, model output, and confidence for each automated decision preserves traceability for incident review and compliance.

Structured alert context

Supplying parsed fields such as source IP, hash, and rule ID instead of raw logs improves LLM summarization accuracy and reduces hallucination.

AI threat modeling additions

AI systems extend STRIDE with prompt injection, data poisoning, model theft, and supply chain tampering as distinct threat categories.

Immutable model registry

Each model version is stored with a cryptographic hash and deployments reference the digest, preventing silent substitution of artifacts.

MLOps promotion gates

Automated checks for accuracy, bias, and safety red-team results plus a signed model card must pass before a model reaches production.

Canary model rollout

Deploying a new model to a small traffic fraction with rollback triggers limits the blast radius of regressions or safety failures.

AI decommissioning controls

Retiring a model requires revoking API keys, deleting weights and derived caches, and retaining audit records per the retention policy.

AI CI/CD artifact scanning

Pipeline checks scan dependencies and model artifacts for known vulnerabilities, unsafe serialization, and unverified provenance before release.