# What is AvertGate? AvertGate is a runtime security layer that inspects every **tool call** an AI agent makes, both before and after it runs. It sits inline between the agent and the real world, like a gate. ## Why does it exist? Agents no longer just generate text; they act with real credentials: querying databases, sending email, starting payments, deleting files. When an agent is misdirected, the outcome is not "a bad sentence" but **money moved, data leaked, systems deleted**. There are three core threats: - **Indirect prompt injection:** Hidden instructions inside content the agent *reads* (a support ticket, an email, a web page, a tool output). - **Excessive agency:** An agent holding far broader permissions than its job needs. - **Data exfiltration:** PII or secrets leaving through outputs or tool calls, in one shot or dribbled out in small pieces. ## Which layer does it protect? Existing LLM security products mostly protect the **conversation layer** (prompt ↔ response). AvertGate protects the **action layer**: which tool the agent calls, with which arguments, and what comes back. | Layer | What it sees | Who protects it | | --- | --- | --- | | Conversation | Prompt and model response | Prompt scanners, guardrails | | Action | Tool name, arguments, tool result, session history | AvertGate | ## How it works in one paragraph Your agent calls its tools through AvertGate instead of directly. For each call AvertGate (1) checks the policy, (2) scans the arguments for secrets and PII, (3) optionally consults a semantic AI model, and (4) decides: ALLOW, BLOCK or ESCALATE. If allowed, it runs the tool, scans the returned result for prompt injection, and writes the whole event to a tamper-evident audit log. #### Quickstart → Put your first tool behind AvertGate in five minutes. #### Runtime loop → A call's journey from start to finish. #### Policy file → Define what an agent may do in YAML. #### Audit trail → A SHA-256 chained, tamper-evident event log. > **Business model**: AvertGate is a closed-source, enterprise product. After a demo/POC conversation, customers get a private tenant and sign in through an invite link. There is no public signup.