Glossary
Tainted input
Untrusted content such as inbound emails or web pages, which can inform an agent but never authorise an Action.
Tainted input is any untrusted content — inbound email, web pages, retrieved documents, connector metadata, third-party tool responses. One3 tags it at the point of entry. It may propose a value but can never authorise an R3 or R4 Action.
Agents use a privileged planner / quarantined reader split: the model that reads untrusted content cannot invoke risky Actions; it returns structured data to a planner that can. High-risk arguments travel with their provenance, so a bank account read from an invoice email is only evidence for a proposed change, which then needs the approved path and a human. This is One3's main defence against prompt injection.
Related terms
See how One3 would run your business
Tell us about your team and the outcomes you want. We will walk you through the apps, the Digital Employees and the plan for your industry.