Skip to main content
BotGauge attacks your AI agent the way a real attacker would, then tells you what broke, how sure it is, and what to fix. You can go from zero to a first report in about 15 minutes.

​
What an attack means

BotGauge sends your agent carefully crafted messages that try to make it:
  • reveal its hidden instructions (the system prompt)
  • leak customer or internal data
  • drop its rules because someone asked cleverly or hid the request in a document
  • take actions nobody asked for, such as a wrong refund amount, a repeated payment, or touching another customer’s account
  • claim it did something it did not do

​
What you get

For every rule tested, you get a result: broke, held, suspicious, or could not be tested, with the reason. For each problem you see the exact prompt, the agent’s reply, and the tool calls behind it. You can re-check a fix by replaying the attack, and you can view the results mapped to frameworks such as the EU AI Act, ISO 42001, SOC 2, GDPR, and OWASP.

​
The journey

1

Connect your agent

Tell BotGauge how to send your agent a message and read its reply, then test the connection.
2

Let BotGauge learn it

BotGauge chats with your agent to learn what it does. This discovery step is called recon.
3

Review the rules and attacks

See the rules and plan that will run.
4

Approve and start

Nothing attacks your agent before you approve a plan. See Approve and run.
5

Watch, or stop at any time

Results arrive live, and you can stop the run whenever you want.
6

Read the results

Understand the seven possible results and the evidence behind each.
7

Fix, verify, and use the compliance view

Replay a finding to confirm a fix, then open the compliance view.

​
Before you start

Have these ready.
  • A BotGauge account and workspace.
  • A test or staging copy of your agent. Use production only if you must. Production runs read-only attacks only, which limits what can be tested.
  • The agent’s URL and the request format it expects.
  • Any API key or token it needs. BotGauge stores these as secrets and never shows them back.
  • Optional: the agent’s system prompt and a list of its tools, which improve the results.
  • Optional: a developer who can add the BotGauge wrapper to your agent for deeper testing. See black box and grey box.
No agent yet? Use the sample refund agent in the red teaming quickstart to see everything first.

Explore red teaming in BotGauge

Follow the steps from agent connection to compliance report in BotGauge.

Try BotGauge →