AI safety

# Safety you can check.

What stops a mistake before a call, and what records it after.

Principles

## What we build to.

#### Wider, not riskier

AI should let a company do more without adding a risk nobody chose.

#### Built in

Limits and records are part of how each product runs.

#### Visible

You see what your AI called, did and cost.

Safeguards

## What runs today, with its limits.

#### Guardrails on shaped identifiers

Email addresses, card numbers, IBANs, UK National Insurance and US Social Security numbers and API keys are masked or refused. Names and prose are not found.

#### Tools fixed when a workflow is saved

A workflow step calls the tool chosen at save. A visitor’s message fills in values and cannot pick another tool.

#### Tools that change data start off

Only tools that declare themselves read-only switch on by themselves. A person ticks the rest.

#### Untrusted content is marked

A run that read the open web or an outside tool is marked, so its output is checked first.

#### Test runs send nothing

An agent or workflow test run records what it would send, and sends nothing.

#### A second check before acting

A separate model checks an action against your policy. Anything short of a clear approve goes to a person.

#### Opt-outs are honoured

A STOP reply puts that person on the opt-out list, which every campaign send checks.

#### Limits before a call

Key budgets, team caps and a balance hold are checked before a call. A widget has a daily cap and a per-visitor limit.

#### Keys locked to your network

A key can carry a list of addresses, and it is refused from anywhere else.

#### Deny by default

Every route needs an explicit permission rule. A route without one is refused, and the build fails.

#### Widget runs keep no visitor text

A widget run stores no visitor words, and the run is deleted after 30 days.

#### Traces and an audit log

Every routed attempt and every run step is recorded with its cost. Changes to keys, members, teams and agents are logged.

Before and after

## Checked before a call, recorded after it.

Nine of the safeguards above act before a call or an action, and three keep the record afterwards.

**BEFORE A CALL**

Deny by default

Keys locked to your network

Limits before a call

Guardrails on shaped identifiers

Tools that write start off

A second check before acting

Tools fixed at save

Test runs send nothing

STOP replies are honoured

**THE CALL**A model call or an action

**AFTER IT, ON RECORD**

Every attempt and step, with its cost

An audit log of changes

Untrusted content is marked

Widget runs: no visitor text

Press enter or space to select a node. You can then use the arrow keys to move the node around. Press delete to remove it and escape to cancel.

Press enter or space to select an edge. You can then press delete to remove it or escape to cancel.

Good to know

## Where the detail is.

[Trust and securityWhere prompts go, what is stored and what redaction finds.](https://vatan.one/trust)[Privacy noticeWhat Vatan holds, where it lives and for how long.](https://vatan.one/legal/privacy)[Sub-processorsEvery vendor that can receive your content.](https://vatan.one/subprocessors)[Report a security issueHow to report a vulnerability.](https://vatan.one/legal/security)

Report a safety concern

## Tell us what went wrong.

NameWork email

Company, if anyYou are aUser

What happenedInclude the product, how serious it is and who was affected.What you send is used only to respond to this report. See the [privacy notice](https://vatan.one/legal/privacy).

Send the report

## Build it with the limits on.

Tell us what AI should do, and we show you its controls.

[Get in touch](https://vatan.one/contact?reason=security)[Trust and security](https://vatan.one/trust)

Source: https://vatan.one/ai-safety
