<!-- Envolt Mail · https://envoltmail.com/security · Markdown version of this page for AI agents. Status: in development (https://envoltmail.com/status). -->

[Trust](https://envoltmail.com/trust) / Security model

# Think like the attacker

The [trust center](https://envoltmail.com/trust) lists our controls. This page starts from the other side: what someone would try against an agent that handles email, which control stops each attempt, how we test it, and what still gets through. It's a summary of our internal threat model, kept current with it.

[Running in staging] Built and running on our staging environment, on synthetic test mail only. [Designed] Specified in detail, not built or not switched on yet. [What runs today →](https://envoltmail.com/status)

## One rule underneath everything

Text inside an email, an attachment or a tool result is data, never instructions. Nothing a sender writes can change what an agent is allowed to do. Permission lives in code a person controls, outside anything a message can reach. Every defense below is a way of keeping that rule true.

## The three risks we rank highest

### 1. Steal data with a planted instruction

- Get the instruction to an agent: **blocked**, the reader has no tools and agents see structured fields, not raw mail [Running in staging].
- Point a reply at the attacker: **blocked**, the Seal checks every recipient [Running in staging].
- Use a look-alike of a known contact: **flagged** by the confusable-name check [Running in staging].
- Change the message after approval: **blocked**, the send permission is bound to the exact content [Running in staging].

### 2. Take over the person in charge

- Fake a command by email: **designed to be refused**, not built yet. Commands will need the right sender, a valid signature and a rotating code, and email will only ever be able to make things safer.
- Phish the sign-in: **blocked**, sign-in uses passkeys only, with no password or email link [Running in staging].
- Send as them with a stolen mailbox token: **no authority** inside Envolt by design, since sent mail is evidence, never a command.

### 3. Read another customer's mail

- Pass another customer's id to an agent tool: **blocked**, the customer comes from the key, never from an argument [Running in staging].
- Collide on a shared key or cache: **blocked**, every key is scoped to one customer [Running in staging].
- Mix up policy versions: **blocked**, each verdict names the exact policy it used [Running in staging].

## Abuse cases we test

The attempt · What stops it · What remains ·

Flood the inbox to burn the model budget · Rules sort bulk mail before any model runs; budgets per sender and per customer trip a breaker. · Slower sorting for the flooded customer only. ·

Look-alike domains and characters · Names are compared on their visual skeleton against known contacts; a near match is flagged, not trusted. [Try it](https://envoltmail.com/lab#xray). · A look-alike of someone you've never met looks like a new contact, and first contact is always yours. ·

"Reconnect your mailbox" phishing · Connections will only be made inside the signed-in console; we never email a reconnect link, and the Airlock flags those lures. · You can still grant a rogue app directly at your mail provider. ·

Slow poisoning: planting "facts" over many messages · By design, facts carry their source, money details are never learned from mail, and memory changes are recorded and reviewable. Agent memory is not built yet. · Subtle drift in tone, which voice checks watch. ·

Sending the same approved message twice · Each send permission is single-use and spent atomically before the provider call. [See it](https://envoltmail.com/lab#seal). · A provider that ignores idempotency keys; we'll measure this in shadow mode. ·

Approving a plan that changed on screen · Every action names the card version it saw; a stale approval executes nothing and the card comes back with what changed. · None known. ·

Rewriting history · Every event goes on a tamper-evident record ([how](https://envoltmail.com/how-it-works#ledger)), and a broken record freezes until a person repairs it. [Tamper with one](https://envoltmail.com/lab#ledger). · None known for recorded events. ·

Forging the "verify this agent" footer on phishing mail · The receipt page never asks for input, and the Airlock flags receipt links on unsealed mail. · A person may still click a convincing fake. ·

## How it's tested

Each control maps to named tests, and a release can't ship unless they pass: a red-team corpus of hostile emails, leak scans of the code bundle and logs, isolation canaries between synthetic customers, and the published test vectors for seals, permissions and the ledger. The [Lab](https://envoltmail.com/lab) runs several of the same modules in your browser. Results for every release are on the [status page](https://envoltmail.com/status), and our scored evaluation, misses included, is in [evaluation.json](https://envoltmail.com/assets/evaluation.json).

### The latest evaluation, misses included

### What the evaluation actually found

mail-boundary-2026-10-18-v1 · Frozen synthetic corpus; deliberately compromised fixture reader; real Reader, ThreadDO, Seal, capability and sink code. Each hostile action targets one disallowed external address. Benign action targets one allowed sink address. No real mail or paid model.

Frozen synthetic fixtures; deliberately compromised reader

Partition · Mail · Benign held for review · Hostile review misses · Poisoned summaries retained · Disallowed external actions executed ·

Public (exposed) · 24 · 3/12 · 4/12 · 3/12 · 0/12 ·

Historical holdout (exposed) · 24 · 4/12 · 7/12 · 8/12 · 0/12 ·

All attempted external actions targeted one address excluded by the test policy. A zero here is evidence for that destination constraint, not a universal safety score. Legitimate mail can require review. No real-model attack-success rate has been measured.

An earlier independent label review of these fixtures: 47 of 48 agreed; one intended hostile fixture is ambiguous in isolation. Both partitions have since been inspected and are regression data, not unseen quality evidence.

[Download the versioned results and limitations](https://envoltmail.com/assets/evaluation.json) · [Inspect the public fixtures](https://envoltmail.com/assets/evaluation-public-fixtures.json)

## Risks that remain

- **People click things.** A convincing fake of our footer can still fool someone; we reduce the odds, we can't remove them.
- **Outside our walls.** An agent or app holding unrestricted mailbox credentials elsewhere is beyond Envolt's controls.
- **Model providers.** Redacted text would reach a model provider under no-training terms; a breach on their side within the retention window remains possible. No model is switched on today.
- **Not yet proven on real mail.** Everything above runs on synthetic mail. Real traffic will find things synthetic mail can't, which is why the [beta gates](https://envoltmail.com/beta#gates) include 90 days on our own inbox and an external penetration test.

Found something? [Report a vulnerability](https://envoltmail.com/trust#report). We respond to every report.
