GigaMail IT中文 GitHub Download

Security

What the agent can do, what it cannot, and what is still open.

GigaMail hands an AI agent access to a mailbox. This page lists every protection that is in place today, the limits we know about, and what we plan to build next. It grows with each release.

The agent proposes. You decide.

An agent can ask to send a message. It never receives anything it could use to approve its own request.

01 · PREVIEW

The agent gets a preview

A send, reply or delete returns the preview and an inert request id. The agent cannot turn that id into an action.

request req_7f3a · inert
02 · APPROVAL

You approve, outside its reach

Console, command line, Telegram or a Windows notification, confirmed with Windows Hello or Touch ID.

Held · awaiting approval
03 · EXECUTION

Exactly what you saw runs

The arguments you reviewed, never a replacement. Every write lands in an append-only action log.

Approved by you · sent

Ten tools wait behind this gate: sending and replying, moving and deleting messages, deleting folders, creating and deleting calendar events, creating Zoom meetings, uploading and deleting Drive files. The MCP map lists all 31 tools with their risk class.

Bound to what you saw

An approved request runs with the account, recipients, folder and attachments frozen when it was created. Attachments are bound to their SHA-256 and size: a file swapped on disk after your approval stops the send.

It fails closed

If the approval store cannot be read, nothing runs: the answer is a named refusal, not an error someone might later silence. On a machine with no way to verify a person, approval is refused rather than replaced by a yes/no question.

Who can approve, strongest first

ChannelHow it worksHow strong
Windows Hello / Touch IDAsked at every approval, on the device. A program can open the prompt; only the person at the machine can satisfy it. On by default on Windows and macOS.STRONGEST
Local PIN on LinuxTyped in an interactive terminal, stored only as a hash, locked for 15 minutes after 3 wrong attempts. An agent running the command from a script gets no prompt.WEAKER
TelegramOpt-in. A tap from the chat you registered behind the OS prompt. It works from any device signed into your Telegram account, including a desktop client left open.WEAKEST

Rejecting never needs any of these: saying no is always safe. If you enable Telegram approval, do not keep Telegram signed in on the machine the agent uses.

Automations, only if you sign them

The one way a mail leaves without a click: a rule you create in the app, for senders and documents you choose. Semi-auto still asks each time; auto runs inside the scope you signed.

No tool can switch one on

A rule is created and resumed only behind Windows Hello or Touch ID, from the app or the command line. No MCP tool can create or change a rule, so a mail cannot switch on autopilot. Pausing needs no prompt.

Narrow by design

Specific senders or one folder, only the documents attached to the rule, a daily cap, a pause between replies to the same sender, and a mandatory expiry. Never “everything”.

The agent writes the body, nothing else

Recipient, subject and thread come from the incoming message. The reply goes to the sender that matched the rule, never to an address written in the mail or in the draft.

Rules decide whether to reply, not a model

No automatic reply to mail that fails sender authentication (DMARC), to automatic mail, mailing lists, no-reply senders, spam, or mail with executable or archive attachments. The first message from a new sender goes through you. A burst of matches pauses the rule.

Every automatic send is written to the action log with the rule that let it through, and you get the same notification as for an approval.

Your data stays local

No GigaMail server

Index, archive, credentials, memory and configuration live on your machine. GigaMail runs no servers and receives nothing. What the agent reads is handled by that agent, under its provider's policy.

Credentials never pass through the agent

Accounts are added by you, in the app or in your own shell. There is no tool an agent could use to add an account or read a password, a token or a key.

One account at a time

Every tool names the account it works on. An agent asking for one mailbox cannot read or write another.

Only the files you registered

The agent can quote only the documents you added as knowledge files. Paths outside that list are refused, links and shortcuts included.

DESKTOP APPSensitive fields hidden

The masker replaces tax codes, VAT numbers, IBANs and phone numbers before a mail goes to the agent for a draft, and puts them back in the draft. Rules and checksums, no AI. The same filter for MCP agents is planned.

An email is input, not authority

Message content is treated as untrusted data. An instruction hidden in a mail cannot add an account, read a secret or approve a send.

The agent reads what you see

Text a person cannot see (hidden blocks, zero-size fonts, white on white, invisible characters) is left out and flagged. The agent gets the visible text, never the raw HTML. A mail that cannot be checked is withheld, not guessed.

Programs never reach the agent

An executable attachment is not downloaded for an agent, is marked with a warning, and is replaced by a note in the archive. The desktop app never opens one; you can still save it yourself.

No link opens by itself

Text written by an agent is shown without loading images or previews, in the app and on Telegram. A link that carries data cannot be fetched without a click.

Tested against hostile mail

Anti-injection tests run on every change, and a red-team run feeds hostile emails to a real agent with every tool enabled, behind a guard that records destructive actions instead of running them.

The desktop app treats mail as hostile ground

Mail cannot run code

Every message passes through one sanitizer and is shown in a sandbox where no script runs. Links open only as web or mail addresses.

Local and token-gated

The app talks only to its own backend on your machine, with a random token per installation.

Signed releases

Built from the tagged source. Windows installer and app are signed; the macOS app is signed and notarized. Linux packages are not signed. SHA-256 digests are on the release page.

Connections verified

Mail server certificates are checked by default. A server with a self-signed certificate is accepted only for the account where you allowed it, behind Windows Hello or Touch ID.

What this does not cover

  • What your agent decides to ask. A hostile mail can still persuade an agent to request an unwise send. The request stops at the preview: that is the design working, and the preview is where to look.
  • An agent with a shell that goes around GigaMail. If it can run any command as you, it can drive your mail program or your provider directly. No mail gate can prevent that. Keep the agent's own command approvals on.
  • Your model provider. What the agent reads leaves GigaMail's boundary and follows that provider's data policy.
  • The action log is append-only, not tamper-proof. Anyone who can write to your disk can alter it.
  • Visible instructions. Text a person can read still reaches the model. Hidden-text removal narrows the gap between what you see and what the agent reads; it does not close it.
  • A machine or user account that is already compromised.

Planned

WhatWhyStatus
The masker for MCP agentsToday sensitive fields are hidden only when the desktop app asks the agent for a draft. The same filter will apply to agents connected over MCP.PLANNED
Telegram approval from the phone onlyToday a tap from any signed-in Telegram client approves. Approval will move to a Telegram Mini App that asks for the phone's biometrics, so desktop and web clients cannot approve.PLANNED
Stricter hidden-text readingInherited styles, more colour forms and font sizes, and complex style rules that can only hide text.NEXT RELEASE

Found and fixed

ReleaseWhat was wrongWhat changed
0.4.1The mail server's certificate was not verified on IMAP connections.Certificate and host name are verified by default; a rejected certificate stops the connection.
0.4.0An approved attachment could be swapped on disk before sending.Attachments are bound to their SHA-256 and size; a mismatch stops the send.
0.1.4An agent with a shell could run the approval command itself.Approving requires Windows Hello or Touch ID, from every path.
0.1.1The agent received a confirmation token and could approve its own request.Approval moved outside the agent's reach; it receives only an inert request id.

Some of these were reported by people reading the code. They are credited in the full security policy.

Report a vulnerability

Please do not open a public issue. Use GitHub private vulnerability reporting, or write to founder@adecubed.com with “GigaMail security” in the subject. You get an acknowledgement within 72 hours and an assessment within 7 days. There is no bounty programme: there is a public thank-you in the release notes, unless you prefer to stay anonymous.

Full security policy on GitHub