The AI agent firewall for the attacks that fool people and AI.

An AI agent firewall checks what an AI agent is told, what it tries to do and what it says, and stops the turns that are unsafe. Autessa AI Agent Firewall inspects the user’s message, every tool call, every tool result, the final reply and the caller’s voice, and can flag, ask the person to verify, hold for an admin’s approval or block.

Yes. Turn it on and every Autessa agent is covered by the company policy, starting in Monitor so you can see what it catches before anything is blocked. Then shape it to how your company works: per-agent policies, your own rules and exceptions, which tools need verification, who sees the word of the day and who gets alerts.

Yes. The firewall scores user messages and tool results for prompt injection. A flagged injection also makes every later tool call in the same run that sends data to another site wait for approval, which is how indirect prompt injection usually tries to exfiltrate data.

Yes. Every caller utterance longer than about a second is scored from 0 to 1 for how likely it is to be synthetic, labeled Real, Synthetic or Unsure. At 0.5 or above your chosen action applies. The score is a signal, not proof, so pair it with verification for anything sensitive.

It is a short word that changes every midnight and is visible only to your team, in the web app and the mobile app. Before a sensitive tool runs, a voice agent asks the caller to say a three-word phrase with today’s word in the middle, so someone replaying a recording or guessing cannot pass.

The agent responds exactly as if verification passed, so the person applying pressure cannot tell. Behind the scenes every tool that needs verification quietly answers that it is unavailable, no sensitive records are shared for the rest of the conversation, and admins and any named contacts get an urgent alert.

No. The word, passphrase or code goes straight to the check. In the conversation history it is replaced with “[verification answer hidden]” and never reaches the model, your logs or stored transcripts. On calls, authenticator codes are typed on the keypad so the AI never hears them.

Yes. The AI Application Firewall API takes each turn of your own conversation, at the user message, tool call, tool result, reply or voice phase, and returns ALLOW, FLAG, CHALLENGE, REQUIRE_APPROVAL, BLOCK or DURESS with reasons. Named policies keep different settings per application.

No. New policies start in Monitor, which records what each check would do without changing anything your customers see. Read the alerts and dashboard for a week or two, tune the thresholds, then switch to Enforce. The attack simulator lets you test before real traffic.

No. The voice check scores each utterance and keeps only the score. It never builds or stores a profile of anyone’s voice. Alerts, scores and verification records are deleted after 30 days, and unusual-request history is kept as counts and summaries, not copies of messages.

It is charged per check that actually runs, at the same price for Autessa agents and the API. A check that is on but does not apply, such as the caller ID check in a chat, costs nothing. Talk to us for pricing on your plan.