What is an AI firewall?
An AI firewall is an inline enforcement point that inspects requests made by AI agents — API calls, tool invocations, data access and agent-to-agent messages — and allows, limits or blocks them based on identity, capability, policy and risk. Unlike a network firewall, it reasons about who the agent is and what it is trying to do, not just packets and ports.
- Inspects AI identity, APIs, tools, agent-to-agent traffic and data access.
- Fails safe: offline deny/limit rules when the control plane is unreachable.
- Quarantine and revocation propagate to every enforcement point.
How it works
- Intercept — Requests from agents pass through the firewall as a proxy, SDK hook or gateway.
- Identify — The caller’s workload identity is verified; unattested callers are blocked.
- Evaluate — Capability, communication, data-access and agent-to-agent policies are applied with a live risk score.
- Enforce — The request is allowed, limited (scope, volume, rate) or blocked; anomalies can trigger quarantine.
Frequently asked questions
How is an AI firewall different from a web application firewall (WAF)?
A WAF protects applications from malicious HTTP traffic. An AI firewall governs what autonomous agents are authorized to do — tool use, delegation, data access — using identity and capability-based policy.
Does the AI Firewall stop prompt injection?
It limits the impact. Even if an agent is manipulated, it cannot use capabilities it was not granted, and high-impact actions still require out-of-band authorization.
What happens if the firewall loses connection to ControlCore?
It keeps enforcing from a signed local policy cache and applies offline deny/limit rules. It does not fail open.