How ChaosDesk keeps AI answers safe

When AI answers customers, people will try to trick it. ChaosDesk assumes the model can be fooled and checks everything around it.

What goes in

• Customer text is treated as data, never as instructions. Hidden characters and fake instruction tags are removed.
• Articles and macros used as sources are cleaned of hidden comments and images.
• Personal data is masked before anything is sent to the AI provider.
• Messages that look like an attack (“ignore your rules…”) are handed to a person without calling the AI at all.

What comes out

Every answer passes a filter that:

• blocks answers that repeat the AI’s instructions or contain secrets such as API keys;
• removes images, HTML and dangerous links. Links are only allowed to your help centre and your sources;
• masks personal data, unless it came from your articles or the customer’s own message;
• keeps answers short.

If the filter blocks an answer, the conversation goes to a person.

Actions

The AI agent can only use the HTTP actions you set up, with the parameters you declared, over https, on the action’s own host. Actions that change data need the customer’s explicit confirmation in the conversation.

Limits

Each visitor can only ask a limited number of questions per 10 minutes and per day, and each workspace has hourly and daily limits on AI use. Any limit hands the conversation to a person, so a script can’t run up your bill.

What the AI won’t do

It answers only from your articles and procedures. When they don’t answer the question, it hands over instead of guessing. It doesn’t promise refunds or give legal or medical advice.

Related

• AI and your data
• Set up the AI agent