Policy reference

All 54 built-in policy types across five categories, what each one does, and where in the pipeline it belongs.

How policies work

A policy is one processing step attached to a proxy at a specific point in the request pipeline. You attach it, order it, optionally give it a condition, and configure it through a form generated from that policy type’s field schema.

• Category groups policies by intent - it does not restrict where they can be attached.
• Flow decides when it runs: proxy request, target request, target response, proxy response, or on fault.
• Order within a flow is execution order. Verifying an API key before applying that key’s quota is a choice you make here.
• Condition makes a policy apply selectively - one route, one client type, one header value.
• Enabled can be turned off without detaching, so the configuration survives while you debug.

The forms are generated, not hand-built

Every policy type carries a field schema on the backend, and the configuration form is rendered from it. Adding a new policy type is a backend-only change - no frontend release is needed for its form to appear.

Traffic management (11)

Protecting your backend from load, and protecting your latency from your backend.

Policy	What it does	Usual flow
Quota	Caps the number of calls an app or developer may make over a time interval.	Proxy request
Spike Arrest	Smooths traffic bursts by throttling the instantaneous request rate.	Proxy request
Concurrent Rate Limit	Limits simultaneous in-flight connections to a backend target.	Target request
Response Cache	Caches backend responses at the gateway to cut latency and origin load.	Proxy request / response
Lookup Cache	Reads a value from the shared cache into a flow variable.	Proxy request
Populate Cache	Writes a flow variable into the shared cache under a key.	Target response
Invalidate Cache	Evicts one or more cache entries, typically after a write operation.	Target response
Compression	Applies gzip or brotli compression to responses above a size threshold.	Proxy response
Circuit Breaker	Trips open after repeated backend failures and fails fast until the target recovers.	Target request
Retry	Retries idempotent backend calls with backoff on transient failures.	Target request
Timeout	Bounds how long the gateway waits on connect and read for the target.	Target request

Quota vs Spike Arrest

They solve different problems and are usually attached together. Quota is commercial - 10,000 calls a month, tied to the product a consumer subscribed to. Spike Arrest is protective - never more than 50 calls a second, regardless of who is paying.

Security (18)

The largest category, and the one governance standard SEC-002 insists every API draws from.

Authentication and identity

Policy	What it does
Verify API Key	Validates the consumer key presented by a developer app and resolves its API products.
OAuth 2.0	Issues and verifies OAuth 2.0 tokens across the supported grant types.
OpenID Connect	Validates OIDC ID tokens against a discovery document and issuer.
Verify JWT	Verifies a JWT signature, issuer, audience and expiry before the call proceeds.
Generate JWT	Mints a signed JWT for downstream services - last-mile security.
Decode JWT	Decodes a JWT into flow variables without verifying the signature.
Basic Authentication	Encodes or decodes HTTP Basic credentials between client and backend.
SAML Assertion	Validates or generates a SAML 2.0 assertion for federated access.
LDAP / Active Directory	Authenticates the caller against an LDAP or Active Directory tree.
Mutual TLS	Requires and validates a client certificate at the gateway edge.
HMAC Validation	Verifies a request signature computed with a shared secret.

Threat protection and access control

Policy	What it does
IP Access Control	Allows or denies callers by IP address, CIDR range or geography.
CORS	Answers preflight requests and stamps cross-origin headers on responses.
JSON Threat Protection	Rejects malformed or oversized JSON payloads designed to exhaust the parser.
XML Threat Protection	Guards against XML bombs, entity expansion and oversized documents.
Regex Threat Protection	Blocks SQL injection, XSS and other patterns in headers, query and body.
Bot & Abuse Protection	Scores traffic for automated abuse and applies a block, challenge or flag action.
Data Masking	Redacts sensitive fields from logs, traces and analytics captures.

Last-mile security

Verifying a caller at the edge and then forwarding an anonymous request to your backend leaves the backend trusting the network. Generate JWT solves that: the gateway mints a short-lived signed token the backend verifies, so the identity survives the hop.

Mediation (11)

Changing the shape of a message. This is how a modern JSON client talks to a SOAP backend without either side being rewritten.

Policy	What it does
Assign Message	Creates or rewrites the request/response: headers, query params, payload and verb.
Extract Variables	Pulls values out of a message into flow variables for later policies.
URL Rewrite	Rewrites the target path before the call is forwarded upstream.
JSON to XML	Converts a JSON message into XML for SOAP or legacy backends.
XML to JSON	Converts an XML message into JSON for modern consumers.
XSL Transform	Applies an XSLT stylesheet to an XML message.
OpenAPI Validation	Validates requests and responses against the attached OpenAPI specification.
SOAP Message Validation	Validates a SOAP envelope against its WSDL or XSD schema.
GraphQL Protection	Validates GraphQL operations and bounds query depth and complexity.
Raise Fault	Short-circuits the flow and returns a custom error response to the client.
HTTP Modifier	Adjusts status line, headers and status code on the way out.

Extension (7)

Where the built-in types run out, and where the gateway talks to the rest of your estate.

Policy	What it does
JavaScript	Runs a sandboxed JavaScript step for custom logic in the flow.
Service Callout	Calls an external service mid-flow and binds the response to a variable.
Flow Callout	Invokes a reusable shared flow so common logic lives in one place.
Message Logging	Streams structured request and response records to an observability sink.
Key Value Map	Reads and writes encrypted key/value pairs scoped to an environment.
Data Capture	Captures custom dimensions into the analytics pipeline for later reporting.
Distributed Tracing	Propagates W3C trace context and exports spans to a tracing backend.

Extension policies are the expensive ones

A service callout adds a network hop inside your request path, and a JavaScript step adds execution time to every call. Governance standard OBS-001 wants logging or tracing attached; it does not want a callout on a hot path that a cache could have answered.

AI gateway (7)

The module that makes an LLM endpoint a governed API rather than an open door. Every one of these works the same way as the policies above - attached to a flow, configured from a schema, enforced per consumer.

Policy	What it does
LLM Token Quota	Meters prompt and completion tokens per consumer over a billing interval.
Prompt Token Limit	Rejects prompts longer than a token ceiling before they reach the model.
Sanitize User Prompt	Screens inbound prompts for injection, jailbreaks and sensitive data.
Sanitize Model Response	Screens model output for leaked data, unsafe content and hallucinated links.
Semantic Cache Lookup	Returns a cached completion when an incoming prompt is semantically similar.
Semantic Cache Populate	Stores a model completion with its prompt embedding for future reuse.
Model Router	Routes LLM traffic across model providers by cost, latency or weight.

Why an LLM API needs its own policies

• Calls are not the cost unit. Two requests to the same endpoint can differ a hundredfold in price, so a call quota controls nothing. Token quota does.
• The payload is the attack surface. Prompt injection is not caught by a regex threat rule; it needs its own screening step.
• The response is also a risk. Model output can leak training data or fabricate links, so it is screened on the way out as well as on the way in.
• Caching works differently. Two prompts that are not byte-identical can deserve the same answer, which is what semantic caching exploits.
• Providers are interchangeable. A router lets you shift traffic on cost or latency, or fail over, without the consumer changing anything.

Token consumption from these policies surfaces on Traffic & latency and in the ai report type.

Templates: configure once, attach many

A built-in policy type is a blank form. A template is that form already filled in, saved under your own name, and reusable across proxies - “Standard Partner Quota”, “Public API CORS”, “PII Masking”.

• Fork any catalogue entry into a template from the Policies screen.
• Attach it by name from any proxy’s Policies tab.
• Change it in one place when the standard changes.
• Pair it with a required-policy-type governance standard to make it effectively mandatory.