For CIOs and CISOs ready to say yes to AI agents.

The antivirus for AI.

Mountain Theory checks each action your AI asks to take, before it runs. No signatures to update: it checks against your rules, not a list of known attacks.

Good actions go through. Risky ones wait for a person, and bad ones are blocked. Each decision is logged for your auditor.

A Tier 1 public research university in Texas evaluated Palo Alto's Protect AI, Cisco's DefenseClaw, CrowdStrike and Mountain Theory, and selected Mountain Theory.

30 minutes with our CEO: your plans, a look at it working, your first 30 days.

Works with any AI model or framework you use, in your environment or the cloud.

How it works, in three actions. Each action your AI asks to take goes through Mountain Theory before it runs. Reading a vendor's evidence is allowed, and it runs. Finalizing a control assessment is held, and a person decides. Deleting a vendor's evidence file is blocked, and it doesn't run.

Who we are

Founded by Mike May (CEO) and Shaun Cuttill (CTO), security leaders with more than 25 years each.

Our mission is to let society use AI fully and safely, whatever AI becomes. More about us

Optimo AI, our first design partner, tested us side by side: unchecked, the agent deleted evidence files when asked; with us, it was blocked, even when it tried another way. Its founder is one of our advisors, and Optimo is also our business partner.

Patent pending. We first filed in October 2024.

AI with real access is already acting on its own.

  1. Replit, July 2025.

    An AI coding assistant deleted a company's live database after being told not to.

    Source: The Register

  2. November 13th, 2026.

    An FBI warning says attackers use an AI agent platform “to automate multiple phases of cyber intrusions at machine speed.” In one campaign, they took personal information from about 67,000 people.

    Source: FBI FLASH-20261113-001

One was a company's own AI, using access it was given. The other was attackers using AI against other people's systems. Mountain Theory is for the first kind: your own AI, using the access you gave it.

When an AI agent does something it shouldn't, who finds out, when, and what's the record?

  • We check the action before it runs, not after.
  • It's allowed, held for a person, or blocked, by rules you write in plain English.
How Mountain Theory works: an AI agent on any model decides on an action, the action is checked against rules the risk owner wrote in plain English before it runs and is allowed to run, held for a person to answer or blocked, new or unclear cases go to a closer review that also answers allow, hold or block, and each decision is kept in a record; people write the rules and read the record in an admin console outside the action path.

Your auditor gets a record of each decision: who, what, when and why, and how each hold was answered.

What we don't do: no chatbot filter, no rewriting the AI's action.

All 13 things it does, step by step: What is execution-layer security?

A real decision record from the October 21st, 2026 recording: the time, the request to finalize the control assessment, and how the hold was answered: allowed, just this once. The outcome reads ALLOWED.

One held action, start to finish.

  1. Claude Code gets the prompt to finalize the Q3 control assessment in the GRC system, and starts one shell command.

    1 of 5

    Your AI asks.

  2. Claude Code's request to finalize the GRC assessment: the command it asked to run.

    2 of 5

    We hold it.

  3. The Mountain Theory console: the decision on the held action, with answers such as Allowed, just this once.

    3 of 5

    A person answers.

  4. The decided row in the console: the request, allowed, just this once.

    4 of 5

    It runs, once.

  5. The evidence record: the hold resolved, allowed, just this once, and the hold before it, escalated because the rule requires a person's approval.

    5 of 5

    The record keeps it.

Recorded October 21st, 2026: a real session of Anthropic's Claude Code, with Mountain Theory installed from our own installer. The business system it acted on was a stand-in, and a script typed the prompt.

Why it's different

  1. 01

    It checks what AI does, not just what it says.

    Filters read the words. Mountain Theory checks the action.

  2. 02

    It has a third answer.

    Not just yes or no: a risky action waits for a person, only where your rules say.

  3. 03

    It keeps checking when your AI changes.

    In July 2026, after a vendor's model update, our test agent made more than 140 attempts in one run against rules that hadn't changed. None ran, and we changed nothing on our side.

Same AI. Same day. Two endings.

Without Mountain Theory checking, 12 of 12 scenarios were compromised.

With it, 0 of 12.

We ran one AI agent through 12 scenarios, some routine and some deliberately dangerous, side by side: once with nothing checking it, once with Mountain Theory checking each action before it ran.

Compromised means the agent did something its rules didn't allow. Even in a routine vendor review, it finalized a risk rating without the approval its rules required.

More than 440 blocked attempts across three runs, and more than 1,400 decisions. Each decision is an allow, a hold or a block.

It's our own test. The July run is published with its recordings, so you can check it.

Read the July run

Put AI to work. Keep control of what it does.

Bring the AI agent that's still waiting for approval. Tell us what you need to control, and we'll come back with specifics, not a pitch.

Try it on your own AI for 30 days, from $5K.

For your security office: it runs in your environment or the cloud, with a record of each decision, and we fill out HECVAT 4 during the proof of concept.

Higher education is our first market. Running AI on a campus? See it for higher education

Book 30 minutes

From the Threat Lab: The AI Risks You Can Point At Before the Kill Switch Luck and Hindsight More in the Threat Lab

Scroll to Top