> ## Documentation Index
> Fetch the complete documentation index at: https://docs.deepblocker.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Safety and stopping

> How to stop every test at once, what pauses a test by itself, and the escalate-now email for a Critical finding.

A test sends messages to your agent that try to make it misbehave. These are the controls that keep it inside the lines.

## Stop everything

On the **Tests** tab press **Stop all tests**, then **Yes, stop all**. Every scheduled, running or paused test in your workspace stops, DeepBlocker's included. What each test found so far is kept.

Any workspace member can do this. It works even when self-service is off for your workspace. It has its own request limit, so a busy minute never blocks a stop.

## Stop one test

On the test's row press **Stop**, then **Yes, stop it**. A stop is for good; the attacker's working memory for that test is wiped afterwards. To hold a test and come back to it, press **Pause** instead. See [Run tests](https://docs.deepblocker.ai/agent-assurance/run-tests.md).

## DeepBlocker's kill switch

DeepBlocker keeps a kill switch for Agent Assurance as a whole. When it is on, no test anywhere starts or sends a message, and a **Start** press is refused with "Global kill switch is on". DeepBlocker uses it when something is wrong on our side.

## What pauses a test by itself

| Cause | What you see | What happens next |
| - | - | - |
| The month's test conversations are used up | The test is **Paused** with the reason "This workspace has used its N test conversations for the month" | It resumes when the month turns or the plan grows |
| The licence was switched off, or the channel left the plan | The test is **Paused** | Ask your DeepBlocker contact |
| A spending limit was reached | The test is **Paused** with the reason | It resumes by itself when the limit allows, for example on a new day |
| A message failed three times | The message is marked failed | The test may end as **Did not finish**. DeepBlocker is told |

DeepBlocker keeps a spending limit per conversation and per day on the attacker's own cost. Those limits are ours, not yours; they never appear in your report.

## What refuses a message before it is sent

Every message is checked first. Any of these stops it:

* The agent is not proven to be yours, or its address now points to a private network.
* The testing agreement is missing, ended, revoked, or signed for another address or channel.
* The test is paused, stopped or finished.
* The kill switch is on.

See [The testing agreement](https://docs.deepblocker.ai/agent-assurance/agreement.md).

## Escalate now for a Critical finding

When a finding reaches the **Critical** band, a DeepBlocker analyst can send an escalate-now email to the emergency contacts of the agent's testing agreement. For an agreement signed in the console, that is the person who signed it.

The email says:

* the agent and the finding's title,
* the severity level out of 5,
* the rule that was broken,
* the tool to switch off now, or "No single tool is involved; review the agent now",
* what happened, in one paragraph,
* a request to act and reply to confirm.

The escalation is recorded on the agreement. If no contact could be emailed, nothing is recorded and DeepBlocker tries another way.

## Stop conditions in your agreement

The agreement you signed says tests stop at once when the agent shows what looks like real personal data of a real person, when anyone in the workspace presses Stop, or when the agent stops answering normally. See [The testing agreement](https://docs.deepblocker.ai/agent-assurance/agreement.md).

## Where to go next

> [card] [Run tests](https://docs.deepblocker.ai/agent-assurance/run-tests.md)
>
> Pause, resume and stop buttons.

> [card] [Troubleshooting](https://docs.deepblocker.ai/agent-assurance/troubleshooting.md)
>
> Every refusal message and its fix.

> [card] [Support](https://docs.deepblocker.ai/support.md)
>
> Who to write to when something is wrong.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.