Mountain Theory

Mountain Theory is a runtime security control for AI agents that functions like antivirus software for agent behavior, an LLM Guardrails & AI Safety layer. It checks every action an AI agent takes before the action runs, evaluating it against rules written in plain English. Each action is then allowed, held for human review, or blocked. The rules need no signature updates. Every decision is logged for auditors. Mountain Theory works with any AI a company already uses, from coding agents such as Claude Code and OpenAI Codex to models hosted in the company's own environment or in the cloud.

When an agent tries an alternate route around a block, that route is blocked as well. Actions that stray from the task the agent was assigned are flagged. According to the vendor, previously unseen harmful behavior is stopped the first day it appears, without anyone writing a new rule.

Personal data, such as Social Security numbers, health records, and payment details, is stopped before it leaves the company's environment. Instructions hidden inside emails and documents (prompt injection) are blocked at the action they attempt to trigger.

Rules are written in plain English, without code, by the people who own the risk. Each logged decision records who, what, when, and why. Mountain Theory publishes its recorded test results, including the misses.