Visitors sometimes try to push an AI assistant off script: asking it to reveal its instructions, to promise a discount or to repeat a card number. Frosvo guards against this on every answer, and the safety check lets you see it work.
What protects each answer
- Masking. Card numbers, passwords and keys in a visitor's message are masked before the model sees them.
- Warning signs. Messages that look like attempts to override the assistant are noticed and flagged.
- A confidence gate. If the knowledge base does not support an answer, the assistant hands the chat to your team instead of guessing.
- An answer check. Before a reply is sent, Frosvo checks it for leaked instructions and unsafe content. Unsafe answers are held back.
Run the check
- Go to Settings → Websites and find the website.
- Choose Assistant safety check.
- Pick the website if you have more than one and click Run the safety check.
- Wait while it runs. A progress message shows "Running…".
The result reads, for example, "7 of 8 attacks handled safely." Below it, each attack is listed as Passed or Failed, with any warning signs Frosvo noticed in the attack. You see the attack and what a visitor would have seen.
How the check works
Frosvo plants a secret code word in the assistant's instructions and then sends known tricks, such as asking it to ignore its rules, play another role, pose as the system, hide instructions in markup, or repeat a card number. It sends them through the same path a real visitor's message takes, with your knowledge base and your settings.
If something fails
- Read the failing attack and the reply.
- Add the missing facts to your knowledge base, so the assistant answers from your content and hands over when unsure.
- Check the handover rules so uncertain questions reach a person.
- Run the check again.
Run it after you change your assistant's instructions or add many new sources. The results are for your information; they are not a guarantee, so keep an eye on real conversations in the inbox.
Cost
Each run uses a small amount of AI. See AI budget and usage to cap spending. More about protecting accounts and data is in security and privacy.