Glossary

A working glossary of the terms used throughout the Academy: the concepts, the tradecraft, and the vocabulary of AI-assisted offensive security. Treat it as a living reference you keep open while you read the rest of the material.

A

Active Directory (AD) is Microsoft's directory service for managing users, computers and permissions in a Windows network. It is the central nervous system of most corporate environments, and the primary target of internal engagements.

AD CS (Active Directory Certificate Services) is Microsoft's PKI. Misconfigured certificate templates and enrollment endpoints lead to the "ESC" privilege-escalation attacks (ESC1 to ESC11), commonly exploited with Certipy.

Agentic AI describes an LLM given tools and the autonomy to plan and act across multiple steps toward a goal. In offensive security it can enumerate, reason and draft at speed. Its risks are catalogued in the OWASP Agentic Top 10.

AS-REP roasting is requesting authentication data for accounts that do not require Kerberos pre-authentication, then cracking it offline.

Assumed breach is a testing model that starts from the premise that the attacker already has a foothold.

B

BFLA (Broken Function-Level Authorization) is being able to call a function or action you should not, for example an admin operation. A top API risk.

BOLA (Broken Object-Level Authorization) is being able to access another user's object by referencing its id (the API form of IDOR). The number one API risk.

BloodHound is a tool that models Active Directory as a graph and computes the shortest path to Domain Admin.

C

Certipy is a tool for enumerating and exploiting AD CS misconfigurations.

Coercion is forcing a machine (often a Domain Controller) to authenticate to an attacker-chosen host (PetitPotam, PrinterBug, DFSCoerce). The active counterpart to poisoning.

Credential stuffing is replaying username and password pairs from public breaches against a target.

CVSS is the Common Vulnerability Scoring System, a standard 0 to 10 severity score. See the severity rubric.

CWE (Common Weakness Enumeration) is a catalogue of software and hardware weakness types; findings are often tagged with a CWE id.

D

DCSync is abusing replication rights to make a Domain Controller hand over account secrets (including the krbtgt hash).

Defense in depth is layering independent controls so no single failure is fatal.

Domain Controller (DC) is a server running Active Directory. Compromising one (or its machine account) generally means domain compromise.

DPAPI is the Windows Data Protection API; it protects per-user secrets (browser and Wi-Fi passwords) that attackers can recover after compromise.

E

Engagement is one discrete piece of testing work: the scope, evidence, findings and report that belong to it.

ESC1 / ESC8 are AD CS attacks. ESC1 abuses a template that lets a low-privilege user request a certificate as anyone; ESC8 relays a coerced DC to web enrollment for a DC certificate.

Evidence is the subset of artifacts (screenshots, scan output, files) attached to a finding to prove it.

F

Finding is a documented security issue with severity, description, impact, evidence and remediation.

FIDO2 / passkeys are phishing-resistant authentication that cannot be relayed by an adversary-in-the-middle proxy. The durable fix for credential relay attacks.

G

Golden Ticket is a forged Kerberos TGT created with the krbtgt hash, granting arbitrary domain access; a post-compromise persistence and impact demonstration.

Guardrail is a control that stops an AI system from taking a disallowed action, such as an approval gate before credential-class or destructive operations.

H

Hash cracking is recovering a plaintext password from its hash offline, using tools such as hashcat or john.

Human-in-the-loop is a model in which the AI reasons and acts at speed but a human approves every consequential step.

I

IDOR (Insecure Direct Object Reference) is accessing another user's object by changing an identifier.

Impact statement is the part of a finding that says, specifically, who could do what and what they would get. Where risk becomes legible.

J

Jailbreak is a crafted prompt that bypasses an LLM's safety or policy constraints so it produces output it was instructed to withhold. Related to, but distinct from, prompt injection.

K

Kerberoasting is requesting service tickets for SPN accounts and cracking them offline for the service-account password.

Kill chain is the staged model of an intrusion; a lens for where a control would bite.

L

LAPS is Microsoft's Local Administrator Password Solution: unique, rotated local-admin passwords that defeat the "one hash, many hosts" lateral-movement pattern.

Lateral movement is moving from one compromised host to another to expand a foothold.

LLM (Large Language Model) is a model trained to predict text that can reason, summarise and generate. The engine behind AI-assisted testing tools.

LLMNR / NBT-NS are legacy Windows name-resolution protocols with no authentication, abused by Responder.

M

Mass-assignment is binding client-supplied fields straight onto a model, letting an attacker set fields they should not (roles, flags). A recurring API flaw.

MCP (Model Context Protocol) is an open standard for connecting AI models to external tools and data sources through a defined client and server interface. It has its own attack surface: over-broad tool permissions, untrusted tool output, and prompt injection through connected data.

MFA / phishing-resistant MFA is multi-factor auth; the phishing-resistant kind (passkeys) cannot be relayed by an adversary-in-the-middle proxy, unlike OTP or push.

mitm6 is a tool that poisons IPv6 DNS to capture Windows authentication.

MITRE ATT&CK is a knowledge base of real-world adversary techniques; the common language for tagging findings.

N

NetExec (nxc) is the workhorse for spraying credentials, finding where they are admin, and executing or dumping across a subnet.

NTLM relay is forwarding a victim's authentication to a server that trusts them, without knowing the password.

O

OSINT is open-source intelligence: passive reconnaissance from public sources.

OWASP is the Open Worldwide Application Security Project, which publishes the Top 10, the API Security Top 10, the LLM Top 10, the Agentic Top 10, and the WSTG and MASTG testing guides.

P

Pass-the-hash / pass-the-ticket is authenticating with a stolen NTLM hash or Kerberos ticket without cracking it.

Passkey is phishing-resistant authentication; see FIDO2.

Password spraying is trying one common password against many accounts to dodge lockout.

Phishing (static vs AiTM) is credential-harvesting via a fake page (static, where MFA stops it) or a real-time proxy capturing the session token (adversary-in-the-middle, which defeats OTP and push MFA).

Prompt injection is untrusted input that manipulates an LLM into ignoring its instructions or performing unintended actions. Indirect prompt injection hides the instruction in content the model later reads (a web page, a document, a tool result). The number one risk in the OWASP LLM Top 10.

R

RAG (Retrieval-Augmented Generation) is grounding an LLM's answer in documents fetched at query time. It reduces hallucination but expands the attack surface: poisoned or attacker-controlled documents can carry indirect prompt injection into the model.

RBCD (Resource-Based Constrained Delegation) is an AD delegation feature that, when writable by an attacker, allows impersonation of arbitrary users to a target.

Responder is a tool that answers poisoned name-resolution queries to capture or relay Windows authentication.

Rules of Engagement (RoE) are the agreed constraints on a test.

S

Scope is the explicit set of in-bounds and out-of-bounds targets.

Severity is a finding's risk rating (critical, high, medium, low, info), calibrated to risk rather than raw CVSS.

SMB signing is a control that cryptographically binds an SMB session, defeating relay; it is not required by default, which is why relay so often works.

SSRF (Server-Side Request Forgery) is coercing a server into making requests, a cloud-metadata favourite.

T

Threat model is a structured account of what you are protecting, who might attack it, and how, used to prioritise testing and defence.

V

VIEWSTATE is ASP.NET page state, exploitable if it is not encrypted and MAC-validated. Ruling it out (or in) is standard web testing.

W

WAF (Web Application Firewall) is an inline filter for web attacks; it is only a real control if it is in prevention mode and the origin cannot be reached around it.

WSTG / MASTG are OWASP's Web and Mobile Security Testing Guides.