LLM Firewall
The protection around your AI that keeps it safe, on-brand and within budget.
Overview
The LLM Firewall is the layer of protection between your customers and your AI. It screens risky incoming messages before they reach the AI, catches unsafe replies before they are sent, and stops runaway conversations before they use up your credits.
It is on by default for every workspace. Its settings are grouped together in the settings panel: open Agent Builder (or Agent Flow), click Settings and choose Safety & Filtering. Each card there carries an LLM Firewall badge. You can adjust each one, or keep the safe defaults.
Features
Filters risky incoming messages
Message Filtering and Topic Auto-Responses stop spam, abuse, attempts to trick the AI, one-time codes and off-topic messages before they reach the AI. Topic Auto-Responses can also answer common questions instantly with your own templates, translated into the customer's language. You can add your own rules for what to ignore.
Catches unsafe replies before they send
Built-in checks look at every reply for leaked internal instructions, messages that are too long, and promises the AI can't keep. If the AI fails partway through a reply, a safety net sends a polite fallback message in the customer's language instead of going silent, and you are not charged for the failed reply.
Caps runaway conversations
Session Limit & Cooldown caps how many times the AI replies in one conversation per day. If a chat goes in circles, or another bot messages your account, the AI takes a break and resumes later, so one conversation can't quietly drain your credits.
Hands off to your team when needed
With Auto-Pause Duration, the AI waits after your team replies and then resumes by itself, so people and the AI don't talk over each other. When a case needs a person, the AI can hand it over and alert your team.
Tips
Frequently asked questions
Was this page helpful?