ModMailbeta

Security & privacy

An AI in your inbox has to be boring about data.

An email assistant sits on the most sensitive corpus a business owns. ModMail is designed so that giving it access is a bounded decision: what it can reach, how long anything is held, and what it may do without you are all explicit — and the defaults are the cautious ones.

01

A short-lived cache, not an archive

ModMail proxies IMAP and SMTP. Message bodies are held only in a temporary cache that expires minutes after you stop using the tab. The durable copy stays with your provider.

02

Secrets encrypted at rest

Mailbox passwords, OAuth refresh tokens and model keys are encrypted with AES-256-GCM under a key that lives in the process environment, never in the database. A dump of the tables on its own is useless.

03

The agent cannot act alone

Every tool call is authorised server-side against your permission level, and the model cannot talk its way past it. Sending, replying and forwarding need your approval in the UI by default. The one exception is an agent you explicitly set to Acts: it may send inside the caps you gave it, every send is logged, and turning it back to Drafts stops it immediately.

04

No trackers in the product

The app ships no third-party analytics or advertising scripts on authenticated pages, and remote images in message HTML are served through a sandboxed frame rather than fetched with your session.

The permission model, precisely

Each user has one of two levels. Read lets the assistant see the message you have open and answer questions about it; it can draft, but it cannot touch the mailbox. Act adds the tools — labelling, filing, scheduling, preparing a reply in the composer. In both levels the send action itself is a UI step the model has no access to, and the server enforces that independently of what the model believes its instructions are.

The reason is adversarial, not technical: the most dangerous message in an AI mailbox is not one you wrote badly, it is one a stranger wrote well. A model that can be persuaded by the content it reads is a model that can be persuaded to send money, and an approval gate is the only control that survives that.

What is stored, and for how long

DataWhere it livesLifetime
Message bodiesShort-lived server cacheMinutes after the tab closes
Your mail archiveYour provider (Gmail, Outlook, IMAP)Unchanged — never migrated
Mailbox passwords and OAuth tokensEncrypted in the database (AES-256-GCM)Until you delete the account
Model API keysEncrypted per user, per providerUntil you replace or remove them
Agent run logsDatabase, tied to your userKept so you can audit what ran

Model providers see what you send them

When the assistant answers a question, the relevant part of the open message goes to the provider whose key you configured — OpenAI, Anthropic, Google, OpenRouter, DeepSeek or your own endpoint. That provider’s terms then apply to that request, which is precisely why the key is yours: you choose the vendor, you hold the contract, and you can move to a local model on Ollama, LM Studio or vLLM when a thread must not leave your network at all.

Reporting a problem

Security reports go to team@modernlisim.com. Include the mailbox or account affected and a timestamp; we would rather hear about a suspicious interaction twice than not at all.