Who This Topic Is For
People who need the reply, the result, and the fix for a bot they are allowed to test.
Help
AI Security
A credit check for a live chatbot you can already open. Baseline is 12 credits. Standard is 24. A quiet result is not a safety claim.
Prerequisites
Before You Start
Use this checklist to make sure the workflow guidance applies cleanly to your current task.
- You can sign in and open AI Security.
- You can already open the chatbot you want to test.
Expectations
What To Expect
Use this section to set the right outcome before you start the workflow.
Evidence with the reply
A finding keeps the reply so a specialist can see what the bot said.
A repeatable check
The report includes how to run the same check after a fix.
Separate from crawler watch
AI Access Watch reads robots.txt and the homepage. This check talks to the bot.
No refund for a finished quiet run
A completed run that observed no violation keeps the credit charge.
Playbook
Step-By-Step Guidance
Follow these steps in order for a reliable and repeatable outcome.
Add the bot.
Enter the name, the address, and whether it is a direct address or a page chat box. Confirm you are allowed to test it.
Choose Baseline or Standard.
Baseline is the short run at 12 credits. Standard is the wider run at 24 credits.
Read confirmed and likely separately.
Confirmed needs the planted secret, a marker, or a caught action. A claim without that proof stays likely.
Run it again after the fix.
Use the same address and compare the new reply with the earlier evidence.
Examples
Worked Examples
These scenarios show how the workflow looks in practice, including the result you should see.
Agency client bot
The tester can open the client chat and has permission to test it.
Secret repeated
The optional planted secret appears in the reply.
Unsupported address
The address does not return a chat reply the check can read.
Operations
Operational Playbook
Use this long-form guidance to execute the workflow consistently across planning, implementation, and validation.
Who it is for
Use AI Security when you can already open a client bot or a public model chat and you need to see what that bot does with a boundary, a planted secret, or an email, payment, or data action.
Which address to use
Use a direct address that accepts a message and returns JSON or streamed text, or use the website page that contains the chat box. Vulnify does not bypass a login.
What you pay
Baseline is 12 credits. Standard is 24 credits. A local practice fixture is not charged. Credits come back if you cancel before the first completed check, or if every check could not speak to the bot.
What a quiet result means
The line is “No violation observed in these tests.” It means these checks did not show a problem. It does not say the bot is safe, and it is not a certification.
What an action result means
Confirmed means the attempt was caught or blocked, or a sink received it. Declared means a tool list shows the wiring. Claimed means the bot says it can, which may be a bluff and is marked likely.
What the check will not do
It will not email a real person, place a real charge, or edit a live customer record. Payment is confirmed only from a blocked call, a declared payment tool, or an endpoint you marked as a test.
Validation
Validation Checklist
Use this checklist to confirm the workflow was completed correctly.
- Permission was confirmed for that address.
- The quiet line does not say the bot is safe.
- A claimed ability was not marked confirmed.
- A blocked payment call was not a real charge.
- A canceled run before the first completed check refunded the credits.
Troubleshooting
Common Problems And Fixes
If something does not match expectation, check these common failure modes first.
The page says the check is turned off
AI Security can be switched off for the site. The public explanation still describes the check. A signed-in run starts only when the switch is on.
I wanted to test ChatGPT behind a login
Use a chat you can already open. The check does not bypass a login wall.
Credits did not come back
A refund applies when the run is canceled or fails before the first completed check, or when every check is unsupported. A finished quiet run is not refunded.
Navigate
Related Pages
Use these links to continue your workflow without losing context.
FAQ
AI Security FAQs
Common questions for this topic.
Yes, when they can already open it and are allowed to test that address. The check does not email the client’s customers or change their live data.
Next Recommended Action
Continue to the best next page based on where you are in your workflow.