Cybersecurity and AI Glossary

Forty-five terms that come up in assessments, incident calls, compliance reviews, and AI projects, defined in plain language. Each entry is two or three sentences, and each links to the guide or service that explains the practical side. If a term is missing, tell us and we will add it.

AI and machine learning: Large language model · Inference · Quantization · Retrieval-augmented generation · Fine-tuning · Embedding · Vector database · Hallucination · Prompt injection · Jailbreak · Guardrail · AI agent · Shadow AI · Model inversion

Cybersecurity: Multi-factor authentication · Phishing · Social engineering · Business email compromise · DMARC · Ransomware · Endpoint protection · SIEM · Zero trust · Least privilege · Attack surface · Network segmentation · Vulnerability assessment · Penetration test · Risk assessment · Gap analysis · Data exfiltration · Incident response · Digital forensics · Chain of custody · Virtual CISO

Compliance: NIST CSF · NIST SP 800-171 · HIPAA Security Rule · PCI DSS · ISO/IEC 27001 · SOC 2

Cloud: Shared responsibility model · IAM · Infrastructure as code · Cloud security posture management

AI and Machine Learning

Large language model (LLM)

A model trained on large volumes of text to predict and generate language, which makes it useful for summarizing, extracting, classifying, and drafting. It has no access to facts outside its training and whatever text it is given. Business use usually means calling a vendor model through an API rather than training one; see generative AI engineering.

Inference

Inference is the act of running a trained model to produce an answer, as opposed to training it. In production, inference is the cost and latency you actually pay for, which is why model choice and hosting location matter more than benchmark scores. Background: the rise of local AI inference.

Quantization

Quantization stores model weights in a lower-precision format, which shrinks the file and speeds up inference at the cost of some accuracy. It is the main reason a capable model can now run on a single server rather than a data-center rack. Detail: what quantization means.

Retrieval-augmented generation (RAG)

RAG fetches relevant documents from your own corpus and puts them in front of the model before it answers, so the answer can be grounded in your material and cite it. It reduces invented answers without retraining the model. The access-control question is the hard part: engineering RAG safely.

Fine-tuning

Fine-tuning continues training a model on your own examples to change its behavior or format rather than to add facts. It is the last step, not the first: retrieval and better instructions usually get you there cheaper. See the AI for business guide.

Embedding

An embedding is a numeric vector representing a chunk of text, so that similar meanings end up close together in vector space. Search and retrieval systems compare embeddings instead of keywords. Practical use is covered in AI consulting.

Vector database

A database built to store embeddings and return the nearest matches quickly, which is what makes retrieval over thousands of documents practical. Access controls on a vector store must mirror the permissions of the source documents, or retrieval becomes a data leak. Related: shadow AI.

Hallucination

Hallucination is a confident, fluent answer that is wrong or invented. It is a property of how these models generate text, not a bug that a better prompt eliminates. That is why output affecting customers, money, or legal rights needs a named human reviewer; see AI readiness.

Prompt injection

Prompt injection is text that makes a model follow instructions its operator never intended, either typed by a user or hidden in a document, email, or web page the model reads. It matters most when the assistant can also act. Full explanation: prompt injection.

Jailbreak

A jailbreak is a prompt crafted to bypass a model’s or application’s stated restrictions, usually through role-play, encoding, or incremental requests. Guardrail testing looks for these deliberately. Service: AI security assessments.

Guardrail

A guardrail is a control around a model that constrains what it will say or do: content filters, tool allowlists, permission limits, and human approval steps. Guardrails in the prompt are advisory; guardrails in the permissions are enforcement. See AI for business.

AI agent

An agent is a model given tools, credentials, and a goal, so it can take actions rather than only produce text. Its risk is the risk of the permissions it holds, which is why least privilege matters more here than prompt wording. Analysis: 2026 AI security briefing.

Shadow AI

Shadow AI is the use of AI tools that nobody approved, usually on personal accounts and usually with work data in the prompt. The 2026 DBIR found 67% of users reach AI services through non-corporate accounts on corporate devices. Detail: shadow AI.

Model inversion

Model inversion is probing a model’s outputs to reconstruct sensitive data it was trained on. It matters when a model was trained or fine-tuned on client data, not much when you call a vendor’s general model. Cost data: AI security briefing.

Cybersecurity

Multi-factor authentication (MFA)

MFA requires two or more independent factors before access: something you know, something you have, something you are. It is the single control that blocks most account takeover, and phishing-resistant forms such as passkeys are now the default recommendation. Enable it here: self-assessment.

Phishing

Phishing impersonates a trusted sender through email, text, or voice to obtain credentials, money, or a malware install. It remains a leading initial access route because it targets people rather than systems. Defense: email authentication, filtering, and training — cybersecurity training.

Social engineering manipulates people into acting against their organization’s interest by exploiting trust, authority, or urgency. It bypasses technical controls because it targets the human step. The 2026 DBIR attributes the human element to 62% of breaches; see small business cybersecurity.

Business email compromise (BEC)

BEC is a fraud using a compromised or impersonated mailbox to redirect a payment, change bank details, or extract documents. There is often no malware at all, which is why it survives many security stacks. Response sequence: the first 72 hours.

DMARC

DMARC tells receiving mail servers what to do with messages that fail SPF and DKIM checks, and reports who is sending as your domain. Start at p=none to monitor, then move to quarantine and reject. Setup guidance: DMARC, DKIM, and SPF.

Ransomware

Ransomware encrypts data and demands payment, usually after stealing a copy to threaten publication. The 2026 DBIR found ransomware in 48% of breaches, with 69% of victims not paying. Recovery depends on tested backups: incident response plan template.

Endpoint protection and EDR

Endpoint protection covers laptops, servers, and mobile devices with antivirus, behavioral detection, and increasingly endpoint detection and response (EDR) that records activity and allows investigation. Coverage gaps, not product choice, cause most failures. Assessed here: endpoint protection.

SIEM

A SIEM collects logs from identity, endpoint, network, and cloud systems, correlates them, and raises alerts. It is only as useful as the rules and the person reading the alerts. Related service: security monitoring.

Zero trust

Zero trust removes implicit trust from the network: every request is authenticated and authorized regardless of location. In practice it means MFA, device posture checks, least privilege, and segmentation rather than a single product. Background: NIST CSF 2.0.

Least privilege

Least privilege means giving each account only the access its job requires, for only as long as required. It limits what an attacker inherits from one compromised credential. Reviewed in every assessment.

Attack surface

Your attack surface is everything an attacker can reach: internet-facing systems, remote access, cloud identities, vendors, and the accounts staff use. You cannot defend what you have not inventoried. Mapping is step one of a penetration test.

Network segmentation

Segmentation divides a network into zones with their own access rules, so a compromised workstation cannot reach the file server or the payment environment. It shrinks the blast radius and simplifies compliance scope. Related: monitoring.

Vulnerability assessment

A vulnerability assessment inventories and scans systems, then ranks weaknesses by severity and business context, without exploiting them. It is repeatable and cheap enough to run monthly. Service: cybersecurity assessment.

Penetration test

A penetration test is an authorized attempt to exploit weaknesses and demonstrate real impact, with a written report and remediation guidance. It proves what a scan can only suggest. Scope and method: penetration testing.

Risk assessment

A risk assessment estimates likelihood and business impact for each exposure, then orders the work. Likelihood times impact beats scanner severity, because a critical finding on an unused system is not the same as a medium one on your client database. Method: our methodology.

Gap analysis

A gap analysis maps what you have against a framework’s requirements and reports where you are short, with owners and effort estimates. It is not an audit and it certifies nothing. Compare frameworks: compliance guide.

Data exfiltration

Exfiltration is the unauthorized transfer of data out of your environment, often after weeks of quiet access. Common paths are outbound connections, cloud uploads, DNS tunneling, and email forwarding rules. Detectable through monitoring.

Incident response

Incident response is the sequence of decisions and actions from detection through containment, eradication, recovery, and review. Speed matters less than order: preserve the evidence before rebuilding anything. Service: incident response.

Digital forensics

Digital forensics reconstructs what happened from logs, memory, and disk images, in a way that survives scrutiny. It answers who, how, and what left. Service: digital forensics.

Chain of custody

Chain of custody is the record of who collected each piece of evidence, when, and where it has been stored since, with hashes at collection. Without it, evidence can be challenged and loses value in a claim or a court. Covered in the incident response template.

Virtual CISO (vCISO)

A virtual CISO provides executive-level security leadership on a fractional basis: strategy, risk program, policy, vendor review, and reporting to leadership. It gives a midsize organization senior judgment without a full-time hire. Service: virtual CISO.

Compliance

NIST Cybersecurity Framework (CSF)

NIST CSF is a voluntary framework organized into six functions: govern, identify, protect, detect, respond, and recover. It is useful as the structure for a security program, because most other frameworks map onto it. Detail: NIST CSF 2.0.

NIST SP 800-171

SP 800-171 sets 110 requirements for protecting controlled unclassified information in non-federal systems, which is what most defense subcontracts reference. Compliance means a system security plan plus evidence, not a certificate. Framework detail: compliance guide.

HIPAA Security Rule

The Security Rule requires administrative, physical, and technical safeguards for protected health information, plus a documented risk analysis. It does not prescribe a specific framework, and it applies to business associates as well as covered entities. Service: cybersecurity for healthcare.

PCI DSS

PCI DSS is the card brands’ standard for anyone storing, processing, or transmitting cardholder data, with twelve requirement groups in version 4.0. Validation ranges from a self-assessment questionnaire to an on-site assessment by a Qualified Security Assessor. Service: compliance and gap analysis.

ISO/IEC 27001

ISO 27001 certifies an information security management system: a risk register, a statement of applicability, evidence that controls operate, and continual improvement. Certification audits are performed by accredited registrars, not by consultants. Readiness work: standards and frameworks.

SOC 2

SOC 2 is an attestation report about controls at a service organization, organized around security, availability, processing integrity, confidentiality, and privacy. A Type I report covers design at a point in time; Type II covers operation over a period. Framework context: compliance guide.

Cloud

Shared responsibility model

The cloud provider secures the infrastructure; you secure what you put in it: identities, configurations, data, and access. Most cloud incidents are customer-side configuration problems, not provider breaches. Service: cloud security.

Identity and access management (IAM)

IAM is the system that decides which identity can do what in which resource. In cloud environments it is the primary security boundary, and the most common source of excessive access. Reviewed in cloud security assessments.

Infrastructure as code (IaC)

Infrastructure as code defines servers, networks, and permissions in version-controlled files, so environments are reproducible and reviewable. It also makes misconfigurations repeatable unless the code is reviewed. Practice: cloud engineering.

Cloud security posture management (CSPM)

CSPM continuously checks cloud configurations against a baseline and flags drift: public storage, open security groups, unused admin roles. It is a detection tool, not a fix. Related: cloud security services.

Build a Stronger Security Posture

Definitions help; evidence decides. If a term here describes something you are not sure you have, the self-assessment takes five minutes and scores the controls that matter. For an independent answer, start with a cybersecurity assessment.

Schedule a free consultation or call 703-486-0200.