Security
Security built into the mailbox
An agent's inbox is two risks at once: a store of sensitive mail, and an input channel that anyone on the internet can write to. nofeb is built to contain both.
Per-tenant encryption and crypto-shredding Shipped
Stored messages, parsed views and metadata are sealed with keys derived for each tenant. Deleting a tenant's key material makes its stored data unreadable, so deletion does not depend on finding every copy.
Keys that cannot widen themselves Shipped
API keys bind to an organization, one pod or one inbox and hold only the scopes they are given. A key cannot mint a key with more rights than itself. Secrets are shown once.
Webhooks that fail closed Shipped
A webhook subscription must name its scope. One without a scope is refused, never widened to the whole organization, so a missing field cannot route one customer's events to another's endpoint. Deliveries are signed per Standard Webhooks.
Routing by the envelope Shipped
Mail is delivered to the inbox named by the envelope recipient, not the To: header, which a sender controls. Redelivery of the same message is idempotent.
Mail checked at the door Shipped
nofeb's inbound server verifies SPF, DKIM, ARC and DMARC and scans for spam and viruses before it accepts a message. Unknown recipients, viruses, clear spam and mail failing a DMARC reject policy are refused during delivery, so nothing is bounced to a forged sender, and every message keeps its security verdict.
Inbound tokens you can rotate Shipped
Mail passed on by Cloudflare Email Routing, Postmark, Mailgun, SendGrid or a raw POST is authenticated by your organization's inbound token, which can only deliver to your own inboxes and verified domains. Rotating it stops every earlier token at once.
Sending policy on the server Shipped
Autonomy tiers are enforced where mail is submitted, for every surface. An agent restricted to read-only cannot send by switching from MCP to the API. Organizations created by agents stay restricted to replies until a human sponsor verifies.
Untrusted mail stays marked as untrusted Shipped
Inbound text is offered to models as safe_body: stripped of invisible characters, hidden HTML and forgeable quoted history, and wrapped in a provenance header. It reduces, but does not eliminate, prompt-injection risk.
Injection scored and quarantined Shipped
Every inbound message is scored for prompt injection, with its reasons. Above the threshold you set, it is quarantined for a person to release or delete, and its content is kept out of the delivery event. Each message also carries a sender-trust label built from authentication, your sender lists, look-alike domains and display-name spoofing.
Lists, caps and approvals Shipped
Recipient and sender allowlists and blocklists, caps per inbox and per key, and held sends that only a person with the separate approval scope can release. Every refusal is an event.
An append-only audit log Shipped
Every change through the API, MCP or the console, and every hold, denial and approval the server makes itself, is recorded with its actor, operation, target, outcome and request id. Query it or export it as JSON Lines.
Single sign-on and SCIM Shipped
People sign in to the console and the webmail through your identity provider over OIDC or SAML 2.0, sign-in can be required for your verified domains, and SCIM 2.0 provisions and removes people. Removing someone ends their sessions and disconnects the apps they approved.
Agents without stored API keys Shipped
An agent can authenticate with the identity its platform already gives it: a Microsoft Entra ID or Entra Agent ID token, a Google service account, an Amazon EKS service-account token, a GitHub Actions OIDC token or any OIDC issuer. You trust the issuer and bind an exact subject, plus conditions such as the repository and branch, to a registered agent with a scope ceiling. Each external token is checked against the issuer's published keys, accepted once, and exchanged for a nofeb token that lasts minutes and cannot be refreshed; nothing secret is stored. Deleting the binding or revoking the agent ends its sessions at once, and every exchange is in the audit log.
Signed agent mail Shipped
Mail an agent sends can carry a signature over its sender, recipients, subject and body, checkable against your organization's published keys, so a recipient can verify which registered agent sent it. When the agent signed in through identity federation, the signature also names the external identity (issuer and subject) that acted.
Further controls
Each control carries its status from the engine registry, and this site's build fails if a page describes one as working before the engine does.
| Capability | What it does | Status |
|---|---|---|
| Sender trust on every message | Each inbound message is labelled trusted, known, unknown or suspicious, with the reasons: its authentication results, your sender lists, look-alike domains, display-name spoofing, and whether the inbox has corresponded with the sender before. | Shipped |
| MCP sign-in with OAuth | MCP clients connect by signing in instead of pasting a key: OAuth 2.1 with discovery, dynamic client registration, client ID metadata documents, PKCE, tokens bound to the MCP server, refresh-token rotation and immediate revocation. A person approves the scopes, and optionally one inbox, in the console, and can disconnect the app at any time. API keys keep working beside it. | Shipped |
| Full export | A whole inbox, a folder, a thread or a search as mbox or a zip of EML files, with every message's folders, keywords, metadata and decisions as JSON, built in the background and downloaded through signed links. | Shipped |
| Retention policies and legal holds | Delete messages older than a number of days, or move them to Trash or Archive, per organization, pod, inbox or folder, keeping flagged mail by default, with a dry run that shows what a policy would do. A legal hold blocks every deletion in its scope, including expiry and inbox deletion. | Shipped |
| AI-generated disclosure | Per organization, pod or inbox, mail an agent sends carries AI-Generated and AI-Disclosure headers and an optional footer. Mail a person sends from the console or the webmail never does. | Shipped |
What we publish, and what we do not
This site describes what nofeb does for you. It does not describe how the engine is built: the vendors under it, the storage layout, the models it will use. That is a deliberate policy, enforced by the same build check that labels every feature, because architecture detail helps an attacker more than it helps a customer.
Reporting a vulnerability
Use the contact form and choose "Security report". Please give us a reasonable time to fix an issue before disclosing it, and do not access data that is not yours while testing.