# What it stops

> The guardrails, agent rules, approval-gated removal and kill switches in Holistic AI EndLayer, where each one acts, and what people see.

Holistic AI EndLayer, the Guardian Agent Platform, acts **at the point of action**: a rule checked before a prompt leaves decides what happens, rather than reporting it a minute later.

## The controls

| Control | Where it acts | What it can do |
|---|---|---|
| **Sensitive data in a prompt** | Claude Code, Cursor and Codex CLI; AI websites through the browser extension | monitor, warn, redact, block |
| **AI browser activity** | AI websites, through the browser extension | monitor, warn, block |
| **Sensitive data in downloads** | AI websites, through the browser extension | monitor, warn, block |
| **Prompt injection in uploaded documents** | AI websites, through the browser extension | monitor, warn, block |
| **Prompt injection and other agent risks** | the Guardian plugin in Claude Code, Cursor, Codex CLI and Antigravity | warn, block |
| **Computer use and impersonation** | Claude Code, Cursor and Codex CLI; agentic browsers | warn, block |
| **AI tools and MCP servers** | the agent, after an administrator approves | uninstall the tool; remove the server from its configuration |
| **Websites** | your device management tool, checked by the agent | block in the browser |
| **Kill switches** | the targeted devices | halt a tool, a desktop app's website, an MCP server, a device group or the whole fleet |
| **Cloud agents' tool calls** | Copilot Studio agents and Amazon Bedrock AgentCore gateways | monitor, or refuse when set to enforce |

Most controls are rules in the Governance policy library. Devices pick up a change within a few minutes.

## Guardrails on prompts

The prompt scanner uses **31 built-in detectors**, covering personal identifiers, secrets and API keys, UK, EU and US identifiers, and source code, plus patterns you write yourself. Each detection carries a **confidence score**, so you can reserve Block for confident matches and let a low-confidence match only warn.

**Redact often matters most.** Blocking a whole prompt because one line holds an account number teaches people to work around the control; replacing that line lets the work continue.

| Outcome | What the person sees |
|---|---|
| **Monitor** | Nothing. The decision is still recorded. |
| **Redact** | The work continues, with the sensitive text replaced. |
| **Warn** | The reason, and in the browser a choice whether to continue. |
| **Block** | A refusal with a reason, never a silent failure. |

The text a detector matched stays on the device. Each event records the detector, the action, a count, the confidence and the policy version that decided it, with masked evidence: the prompt with every match masked on the device.

## Prompt injection in documents

A line in an attached document written for the model, such as "ignore your previous instructions", is a prompt the person never typed. The most effective ones are hidden: white text, tiny text, or characters that render as nothing.

The extension hands each attached file to the agent on the same device **before the site receives it**, and the agent weighs two kinds of evidence:

- **Phrasing:** overriding earlier instructions, giving the model a new role, dictating its verdict, sending data elsewhere, or asking it to keep something from the reader.
- **Concealment:** white or vanishing text in Word, invisible text in a PDF, hidden elements in HTML, and runs of zero-width characters.

An instruction that is also hidden is the clearest case; one phrase alone does not reach the bar. The person sees where the instruction is so they can remove it. EndLayer records the rule, the action and the confidence, never the document.

## Agent behaviour

**The Guardian plugin** is placed by policy in Claude Code, Cursor, Codex CLI and Antigravity. It applies 13 controls, including data leakage, prompt injection and jailbreak, each with a warn and a block threshold.

**Computer use and impersonation** has five rules: agents driving a browser, agents controlling the desktop, agentic browsers, agents sending as the person, and agents using the person's own accounts.

## Removal and kill switches

**Removal is approval-gated.** Blocking a tool requests its removal in the console's Inbox. An administrator chooses to uninstall it, leave it, or allow it on one device. The agent then removes it with the person's own permissions.

**A kill switch halts a tool on the targeted devices.** Claude Code, Cursor and Codex CLI refuse every prompt and tool call until it is released; for the Claude and ChatGPT desktop apps, their websites are blocked. Tripping and releasing a switch both need an administrator and a typed reason, and both are audited.

## Planned

- **Approval before an agent acts:** holding an agent's action until a named person decides.
- **Removing blocked skills.**
- **SDKs:** a decision point inside your own applications.

> **Start in Monitor:** Run new rules at Monitor or Warn first. The events show what Block would have stopped, before it interrupts anybody's work.

Source: https://www.endlayer.ai/docs/enforcement · Updated 2026-11-20
