The words, in plain words.
What JWT claims, SCIM, row-level security, ClusterRoles, MCP and the rest mean, in a few sentences each, with the free tool or guide that goes deeper.
Sign-in and tokens
- aud claim (audience) Also: audience claim, JWT audience
-
The claim that names who a token is meant for; an API must reject any token whose aud doesn’t include it.
Without the audience check, a token issued for one API can be replayed against another that trusts the same issuer. Errors such as “jwt audience invalid” mean the value in the token and the value the API expects don’t match.
See also: JWT claims
- exp claim (expiry) Also: expiration time, token expiry
-
The time after which a JWT must no longer be accepted, as seconds since 1 January 1970 UTC.
Short expiry limits how long a stolen token works. APIs usually allow a few seconds of clock skew; anything more hides real problems. A “jwt expired” error means the client should refresh the token, not that the API should accept it.
See also: Refresh token
- JWT (JSON Web Token) Also: JSON Web Token, jot
-
A compact token, defined in RFC 7519, that carries claims about a user or app as JSON, in three base64url parts: header, payload and signature.
Most JWTs are signed, not encrypted: anyone holding one can decode and read its payload. The signature only proves the token wasn’t changed and who issued it, so an API must check the signature, the algorithm, the issuer, the audience and the expiry before it trusts a single claim.
- Free tool JWT Security Inspector
- Guide JWT Claims Explained: iss, sub, aud, exp and Other Common Claims
See also: JWT claims, aud claim (audience), exp claim (expiry)
- JWT claims Also: token claims
-
The name-value pairs in a JWT’s payload: registered ones such as iss, sub, aud, exp and iat, plus any your issuer adds, such as roles or a tenant ID.
Claims are statements, not proof. Trust them only after the signature checks out, and keep secrets out of them, because the payload is readable by anyone who has the token.
- Free tool JWT Security Inspector
- Guide JWT Claims Explained: iss, sub, aud, exp and Other Common Claims
See also: JWT (JSON Web Token), aud claim (audience), exp claim (expiry)
- MFA (multi-factor authentication) Also: 2FA, two-factor authentication, TOTP
-
Asking for a second kind of proof at sign-in, such as a code from an authenticator app, a passkey or an SMS code, on top of a password.
Authenticator apps use TOTP (RFC 6238): a short code worked out from a shared secret and the time, usually six digits that change every 30 seconds. SMS codes are the weakest option, because phone numbers can be hijacked.
- Free tool Auth Provider Picker
See also: Passkeys
- OAuth 2.1 Also: OAuth
-
The update to OAuth 2.0 that folds in years of security advice: PKCE for every authorization code flow, exact redirect URI matching, and no implicit or password grants.
OAuth lets a user grant an app limited access to an API without handing over a password. The Model Context Protocol builds its authorization on OAuth 2.1.
See also: PKCE, OIDC (OpenID Connect), Refresh token, Protected resource metadata
- OIDC (OpenID Connect) Also: OpenID Connect
-
An identity layer on top of OAuth 2.0: besides access tokens, the provider returns an ID token, a JWT that tells the app who signed in.
Providers publish their settings at /.well-known/openid-configuration, so apps can find the endpoints and signing keys. Enterprise single sign-on runs over OIDC or SAML.
- Free tool Auth Provider Picker
See also: SSO (single sign-on), SAML, JWT (JSON Web Token)
- Passkeys Also: WebAuthn, FIDO2
-
Sign-in credentials built on WebAuthn: the private key stays on the user’s device or password manager, and the site only stores a public key.
They resist phishing because each passkey works only on the site it was made for, and there’s no shared secret on the server to steal.
- Free tool Auth Provider Picker
See also: MFA (multi-factor authentication)
- PKCE Also: Proof Key for Code Exchange, pixy
-
Proof Key for Code Exchange (RFC 7636): the app sends a hash of a one-time secret when sign-in starts and the secret itself when it swaps the code for tokens, so a stolen authorization code is useless.
OAuth 2.1 requires it for the authorization code flow, and the MCP authorization spec asks clients to use its S256 method.
See also: OAuth 2.1
- Refresh token
-
A longer-lived credential an app uses to get new access tokens without asking the user to sign in again.
Because it outlives the access token, it’s worth more to an attacker. OAuth 2.1 asks that refresh tokens issued to public clients, such as browser and mobile apps, are either bound to the client or rotated on every use, so a stolen one stops working.
- Guide JWT Expired: What the Error Means and How to Fix It
- Guide MCP OAuth: How to Add Authorization to an MCP Server
See also: exp claim (expiry), OAuth 2.1
- SAML Also: SAML 2.0, Security Assertion Markup Language
-
An XML-based standard for enterprise single sign-on: the company’s identity provider sends your app a signed assertion saying who the user is.
Many enterprise customers expect it alongside OIDC, because their identity provider and IT policies are built around it.
- Free tool Auth Provider Picker
See also: SSO (single sign-on), OIDC (OpenID Connect), SCIM
- SCIM Also: SCIM provisioning, directory sync, System for Cross-domain Identity Management
-
System for Cross-domain Identity Management (RFC 7643 and 7644): a customer’s identity provider creates, updates and deactivates users and groups in your app automatically.
With SSO alone, someone who leaves the company can keep a live session in your app; SCIM deactivates them there too.
- Free tool Auth Provider Picker
See also: SSO (single sign-on)
- SSO (single sign-on) Also: enterprise SSO, single sign-on
-
Signing in once with a company account, at an identity provider such as Okta, Microsoft Entra ID or Google Workspace, to reach many apps.
For a B2B product, “supports SSO” means each customer can connect its own identity provider over SAML or OIDC, so its IT team controls who gets in and can cut access in one place.
See also: SAML, OIDC (OpenID Connect), SCIM, Multi-tenancy
Access control
- ABAC (attribute-based access control) Also: attribute-based access control, policy-based access control
-
Deciding each request from attributes of the user, the resource, the action and the context, such as department, owner, region or time, rather than from roles alone.
It’s more flexible than RBAC and harder to audit. Many products use roles for the broad strokes and a few attribute rules, such as ownership, for the rest.
See also: RBAC (role-based access control)
- Least privilege Also: principle of least privilege
-
Giving every user, service and AI agent only the access its job needs, for only as long as it needs it.
It limits the damage from a stolen credential, a bug or a misled agent. Read-only by default, and a separate step for anything that deletes, pays or sends, is the practical version.
- Free tool Kubernetes RBAC Explainer
- Free tool MCP Security Checker
- Guide MCP Security Best Practices: Risks and Checklists
See also: RBAC (role-based access control), Human-in-the-loop approval
- Multi-tenancy Also: tenant isolation, organizations
-
One app serving many customer organizations, with every query scoped so a customer only ever sees its own data.
The tenant must come from the signed-in session, never from a value the client can change. Row-level security is one way to enforce that in the database itself.
- Free tool RBAC Matrix Builder
- Free tool Supabase RLS Generator
See also: Row-level security (RLS), SSO (single sign-on)
- Permission matrix Also: RBAC matrix, access control matrix, roles and permissions matrix
-
A table with roles down one side and actions on resources across the top, showing which role may do what.
It’s where product, security and engineering agree on access before code is written, and the checklist an auditor reads afterwards.
See also: RBAC (role-based access control)
- RBAC (role-based access control) Also: role-based access control, roles and permissions
-
Granting permissions to roles, such as admin, editor or viewer, and giving each user one or more roles, instead of granting permissions to people one by one.
It’s easy to reason about and audit. It strains when a decision depends on the record itself, such as “only the author can edit”; that’s where attributes or relationships come in.
- Free tool RBAC Matrix Builder
- Guide RBAC vs ABAC: Differences, Examples and When to Use Each
- Guide RBAC Matrix Template: Free Examples for Excel and Sheets
See also: ABAC (attribute-based access control), Permission matrix, Least privilege
Database rules
- Firestore Security Rules Also: Firebase security rules, Firestore rules
-
The rules language that decides, for every read and write from a Firebase client SDK, whether the request may touch a document.
Server code using the Admin SDK skips them, so they protect only what the browser and mobile apps do. Rules that allow every read or write until a date are a common leftover from test mode.
- Free tool Firestore Rules Generator
- Guide Firestore Security Rules Functions: A Cheat Sheet with Examples
See also: Row-level security (RLS)
- RLS policy Also: Postgres policy, Supabase policy
-
One rule on a table, for one command or all of them, whose USING and WITH CHECK expressions decide which existing rows a request sees and which new or changed rows it may write.
A policy whose expression is simply true lets everyone through, which is the same as having no rule at all for that role.
See also: Row-level security (RLS)
- Row-level security (RLS) Also: RLS, Postgres row level security, Supabase RLS
-
A Postgres feature that filters which rows each request may read, add, change or delete, through rules called policies attached to a table.
In Supabase it’s what keeps data private, because the browser talks to the database through the Data API with a key anyone can see. With RLS on and no policies, the API roles get no rows; with RLS off, they get every row the table’s grants allow.
- Free tool Supabase RLS Generator
- Guide What Is RLS in Supabase? Row-Level Security Explained
- Guide How to Enable RLS in Supabase, and What Breaks Next
See also: RLS policy, Supabase publishable key (anon key), Supabase secret key (service role)
- Supabase publishable key (anon key) Also: anon key, Supabase anon key, publishable key
-
The Supabase key that ships in your app’s pages. It’s public by design and reaches only what row-level security and grants allow.
Hiding it doesn’t protect anything; your policies do.
- Free tool Supabase RLS Generator
- Guide Does the Supabase Service Role Bypass RLS?
See also: Row-level security (RLS), Supabase secret key (service role)
- Supabase secret key (service role) Also: service_role key, service role key, secret key
-
The Supabase key that runs as the service_role, which bypasses row-level security. It belongs on a server only, never in a browser or app bundle.
If one leaks, rotate it at once: anyone holding it can read and change every row.
- Free tool Supabase RLS Generator
- Guide Does the Supabase Service Role Bypass RLS?
See also: Supabase publishable key (anon key), Row-level security (RLS)
Kubernetes
- Kubernetes RBAC Also: Role, ClusterRole, RoleBinding, ClusterRoleBinding
-
Kubernetes’ own access control: Roles (in one namespace) and ClusterRoles (cluster-wide) list allowed verbs on resources, and RoleBindings or ClusterRoleBindings grant them to users, groups or service accounts.
Permissions only add up; there are no deny rules. A ClusterRole bound with a RoleBinding applies inside that one namespace, which is the usual way to reuse a role safely.
See also: Kubernetes RBAC verbs, Least privilege
- Kubernetes RBAC verbs Also: RBAC verbs
-
The actions a rule allows: get, list, watch, create, update, patch, delete and deletecollection, plus the special verbs escalate, bind and impersonate.
list and watch return whole objects, so they expose Secrets as fully as get does. escalate, bind and impersonate let a subject gain permissions it doesn’t have, so grant them rarely.
See also: Kubernetes RBAC
MCP and AI agents
- AI agent
-
Software where a language model decides which steps and tools to use to reach a goal, rather than following a fixed script.
That flexibility is the value and the risk: an agent needs its own identity, the least access that works, logs of what it did, and approval for anything it can’t undo.
- Free tool AI Agent Readiness Scorecard
- Guide AI Agent Use Cases: 18 Examples by Team, and How to Pick One
See also: Human-in-the-loop approval, MCP (Model Context Protocol), AI readiness assessment
- AI readiness assessment Also: AI readiness, AI maturity assessment
-
A review of whether an organization is ready to use AI well: its goals, data, systems, security, people and governance, and which workflows to try first.
The useful output isn’t a score but a ranked list of workflows and the gaps to close before each one.
- Free tool AI Agent Readiness Scorecard
- Guide AI Readiness Checklist: 30 Checks Before You Build AI Agents
- Guide AI Readiness Assessment Framework: Areas, Scoring, Examples
See also: AI agent
- Human-in-the-loop approval Also: human approval, HITL
-
An AI agent pausing for a person to approve an action before it happens, such as sending a refund, deleting records or emailing a customer.
Agents can do the reading and drafting on their own; approval is for the steps that cost money, reach customers or can’t be undone.
- Free tool AI Agent Readiness Scorecard
- Guide AI Agent Use Cases: 18 Examples by Team, and How to Pick One
See also: AI agent, Least privilege
- MCP (Model Context Protocol) Also: Model Context Protocol
-
An open protocol that connects AI applications to tools and data: an MCP server offers tools, resources and prompts, and the client inside an AI app calls them for the model.
Servers run locally over standard input and output, or remotely over Streamable HTTP. A remote server that touches user data needs sign-in, which the spec builds on OAuth 2.1.
- Free tool MCP Security Checker
- Guide MCP Security Best Practices: Risks and Checklists
- Guide MCP OAuth: How to Add Authorization to an MCP Server
See also: MCP server, Protected resource metadata, Tool poisoning
- MCP server
-
A program that exposes tools, resources or prompts to AI clients over the Model Context Protocol, such as “search tickets” or “create invoice”.
Each tool is an action an agent can take with the server’s access, so a server deserves the same review as any API: who can call it, with what scopes, and what each tool can change.
See also: MCP (Model Context Protocol), Least privilege
- Prompt injection
-
Instructions planted in content an AI model reads, such as a web page, an email, a document or a tool’s result, that try to override what its user asked for.
No filter catches all of it. Limit what an agent can reach, keep secrets out of its context, and put a person in front of risky actions.
- Guide MCP Security Best Practices: Risks and Checklists
- Guide AI Agent Use Cases: 18 Examples by Team, and How to Pick One
See also: Tool poisoning, Human-in-the-loop approval
- Protected resource metadata Also: RFC 9728, oauth-protected-resource
-
A JSON document (RFC 9728) at /.well-known/oauth-protected-resource where an API, such as a remote MCP server, says which authorization servers issue tokens for it.
MCP clients find it from the WWW-Authenticate header on a 401 response, then discover the authorization server and start OAuth from there.
See also: OAuth 2.1, MCP (Model Context Protocol)
- Tool poisoning
-
Hiding instructions for the model in a tool’s name, description or schema, so an agent that reads the tool list is steered to misuse other tools or leak data.
Review the tool descriptions an agent sees, pin servers you trust, and ask before any tool that deletes, pays, sends or runs code.
See also: Prompt injection, Human-in-the-loop approval