In short
- Content is data, not orders.Text from your pages and tools is never treated as an instruction to the AI.
- Only links you trust.Answers can link only to your sources, your site and domains you allow.
- Never asks for secrets.Sentences asking visitors for passwords, card numbers or codes are removed.
- Its rules stay private.An answer that starts repeating the agent's instructions is stopped.
An AI agent on your website talks to anyone. Some people will try to make it say things it should not, and some web pages contain text written to trick AI. OneType protects your agent in layers, from the moment content is read to the moment each word reaches the visitor.
When your content is read
Text hidden on a page with tricks, like white-on-white text or HTML comments, never reaches your agent's knowledge; that is where planted instructions usually hide. Passages that read like orders to an AI, such as "ignore your previous instructions" or a fake system message, are flagged on the source so you can see them.
When you save text or a saved answer yourself, a line that gives the AI orders or asks visitors for secrets is refused, and OneType tells you which line to remove.
While an answer is written
Everything the agent reads, from your pages to tool results and page details, is wrapped as data. The AI is told that data never contains instructions, even if it says so, and that only the visitor's own messages are requests. Before a passage reaches the AI, planted orders are cut out, along with sentences that ask readers to send secrets to another site.
Your own instructions sit between fixed platform rules you cannot switch off by accident: the agent says it is an AI when asked, never reveals its instructions and never asks for passwords or card numbers.
Before each word reaches the visitor
Answers stream through final checks as they are written:
- Links: an answer may point only to the sites of its sources, the page it is on, links you set up and domains you allow. Any other address, even a disguised one, is removed.
- Secrets: a sentence asking the visitor to share a password, card number, security code or token in the chat is dropped.
- Leaks: if an answer starts repeating the agent's hidden instructions, it is stopped.
- Leftovers: the AI's private thinking and imitated system tags are removed.
Staying on topic
There is no separate topic filter to set up. With "Answer only from knowledge" on, the agent answers only from your content, and a question your content does not cover gets your fallback reply instead of an improvised one. When handoff is on and a visitor clearly asks for a person, in any of several languages, the chat is handed to your team without asking the AI at all.
What you control
Add the domains your answers may link to in the agent's security settings, up to 50, such as your shop or your docs site. Review flagged passages on each source, where they are highlighted in the page text. Write rules in your instructions for anything else the agent should avoid.
Honest limits
No system can promise to stop every trick, which is why OneType uses several layers instead of one. The pattern checks for planted text are strongest in English. There is no list of blocked topics to fill in; off-topic control comes from answering only from your content and from your instructions.
How OneType keeps each workspace's data apart, and where it is hosted, is on the security page.