Skip to content

Security ​

This page collects what Edge does for security across the platform, in one place. It does not replace the pages it links to — Access control and Permissions-Policy are the detailed references for those two topics.

Transport ​

Every host — live, preview, docs, api, mcp, and app — serves HTTPS only, over TLS 1.2 or higher, terminated at the edge. A plain HTTP request is redirected, never served. The load balancer behind the API and MCP server adds a further restriction on top of that same floor: its cipher suites are limited to forward-secrecy-only.

Every response carries a baseline of security headers, unless the distribution's own headers config replaces one:

HeaderBaseline valueA distribution can replace it
strict-transport-securitymax-age=63072000; includeSubDomainsYes
x-content-type-optionsnosniffNo
x-frame-optionsDENYYes
x-permitted-cross-domain-policiesnoneNo
referrer-policystrict-origin-when-cross-originYes
cross-origin-opener-policysame-originNo
cross-origin-resource-policysame-originNo
permissions-policyCamera, microphone, geolocation, payment, USB, and the interest cohort all deniedYes, within limits — see Permissions-Policy

A distribution can also add content-security-policy, access-control-allow-origin, x-robots-tag, and cache-control, none of which has a platform default. The headers marked "No" above are enforced on every response regardless of distribution config and cannot be changed. Set-Cookie cannot be sent through this config either — see Sessions below for how the visitor-access cookie is actually set.

Network abuse protection ​

A web application firewall sits in front of every surface and blocks an address outright once it crosses a limit, independent of access rules or network rules:

SurfaceLimit
Any request to a distribution2,000 per 5 minutes, per address
Sign-in (/_auth)100 per 5 minutes, per address
Inbound hooks (/_hooks)600 per 5 minutes, per address
Signed uploads (/_upload)20 per minute, per address
MCP and the HTTP API2,000 per 5 minutes, per address, plus a managed rule set that blocks common injection and scripting patterns

A heartbeat prober used for Edge's own uptime checks is exempted from the rate limits so a health check never competes with real traffic for the budget.

Visitor and caller access ​

Who can reach a distribution, and who can sign in to one that requires it, is covered in full in Access control: login modes and methods, access rules, IP and country network rules, and sessions. The short version:

  • Network rules (ip_allow, ip_block, country_allow, country_block) are checked before sign-in and apply to every visitor.
  • Access rules (email, email_domain, any_signed_in, deca_member) decide who may sign in once a site requires it.
  • Turning access down — disabling login, dropping require_mfa, or removing a sign-in method — needs permission to change access and is rejected otherwise (ACCESS_DOWNGRADE).

Sessions ​

A signed-in visitor's session is a cookie whose value is HMAC-signed with a key derived per distribution; Edge never stores the session contents server-side beyond what the cookie carries. Signing everyone out rotates that distribution's key, which invalidates every cookie signed with the old one; cookies issued after that point use the new key. Passwords are hashed with Argon2 before storage — Edge never stores or logs a visitor's password in plain text.

Callers: personal access tokens ​

A personal access token authenticates a script, a CI job, or an agent as the person who created it, scoped down from there:

  • It can never hold more than its creator's own role.
  • It is bound to one workspace. A request for another workspace is rejected.
  • It is shown once. Edge stores only a hash of the token, never the token itself, so a database compromise does not expose working credentials.
  • It carries an expiry you choose — 7 to 90 days, a year, a fixed date, or no expiry — and can be revoked at any time, which takes effect immediately.

See Permission scopes for the full list of scopes a token or an agent's JWT can carry, and Organizations and workspaces for how a request is bound to its organization and workspace.

Tenant code isolation ​

A function runtime runs in its own WebAssembly sandbox for each invocation, with no ambient capability beyond what the call explicitly needs:

  • No sockets, no filesystem, no environment variables, and no timers are available inside the sandbox.
  • CPU is metered — a runaway handler exhausts its own budget and stops, without affecting other tenants or other invocations on the same host.
  • A handler also runs under a wall-clock deadline; hitting it ends the call the same way.
  • Outbound HTTPS (network.native) is host-mediated, not a real socket handed to the guest: only port 443, only a DNS name rather than a literal address, no redirects followed, no private, link-local, or metadata-service address, and no visitor or platform credential is ever attached to the call. A function cannot use this path to reach another tenant's infrastructure or an internal address.
  • Reading or writing a data object, a key-value or SQL store, or a file prefix needs an explicit grant on that function; access is denied by default.

Outbound and inbound integrations ​

  • Outbound webhooks are signed per the Standard Webhooks specification (HMAC-SHA256 over the message id, timestamp, and body). Signing secrets are wrapped with a dedicated encryption key; rotating a secret keeps both the old and new signature valid during the overlap, so a receiver is never dropped mid-rotation.
  • Webhook delivery resolves the destination once, confirms the resolved address is public, and then connects to that exact checked address while TLS and the Host header still name the original hostname. Resolving again at connect time would let a hostname answer publicly for the check and privately for the delivery; pinning the checked address closes that gap. A redirect from the destination is not followed.
  • Inbound hooks (/_hooks/{slug}) verify the sending provider's signature before anything downstream — a function, a queue, or a workflow — ever sees the request, and deduplicate by idempotency key so a provider's retry is not processed twice.

See Events, webhooks, and hooks.

Data protection ​

  • Stored content, data objects, and key-value and SQL storage are encrypted at rest with customer-managed encryption keys. Files a visitor uploads through a form are encrypted with a separate key from the one used for distribution content.
  • A visitor-uploaded file is scanned for malware and checked against the type its field declared before it is ever downloadable. A file that fails either check (malware, type_mismatch, or size_mismatch) is deleted, not merely hidden; a file whose scan never reaches a verdict stays permanently unreadable. See Accept file uploads on a form.
  • A generated download link for a submission's file is single-use in the sense that it only works once the file is clean, expires after 7 days, and every fetch through it is logged.
  • PII detectors flag emails, phone numbers, payment-card and government-id numbers, JWTs, and PII-shaped key names in analytics properties and form submissions. A workspace admin resolves each detection as allow, redact, or block; payment-card and government-id detections are locked to block and cannot be allowed through.

Logging and request data ​

  • Request logs include the visitor's IP address and country for security and abuse review, gated behind the logs.read scope. They never include the query string, cookies, or the request body.
  • Analytics never stores the visitor's IP address, regardless of logs.read.
  • Every read of a form submission or an uploaded file is logged, and reads are capped per day — see Let an agent read form submissions.

See also ​