Skip to main content
Agent Assurance tests your AI agent the way a real attacker would. Our attacker talks to your agent over its own channel and tries to make it break a rule. Every attempt is recorded. You get a report with an exposure score and the proof behind each finding.

​
What a test is

A test is one or more conversations between our attacker and your agent. Each conversation has a goal, such as “get another customer’s details”. The conversation runs until the attacker succeeds or gives up, and never past the message limit you set. See Test modes for how deep a test can go.

​
Who does what

Roles come from your workspace. See Team and roles.

​
The switch

Agent Assurance is turned on per workspace by DeepBlocker. Self-service, where you add agents and run tests yourself, is a second switch. Until both are on, the console shows “Your workspace does not run its own tests yet. Ask your DeepBlocker contact to turn this on.”
To get self-service turned on, write to hello@deepblocker.ai with your workspace name. DeepBlocker turns it on and tells you. Nothing else changes in your workspace.
When the switch is off you can still open the console, read reports and stop any test. Stopping always works.

​
What you need

  • An agent that answers on a public web address or as a Telegram bot. See Channels.
  • A plan with test conversations left this month. See Plans and usage.
  • Someone with the owner or admin role to sign the testing agreement.

​
Where to go next

How a test runs

The steps from adding an agent to reading the report.

Add an agent

Add your first agent and prove it is yours.

Connect a web agent

The request we send, the reply we need and the signature to check.