> ## Documentation Index
> Fetch the complete documentation index at: https://docs.hiddenlayer.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Configuration Overview

HiddenLayer Agent Harness Security protects your AI applications while you run them. Security policies are how you tell Agent Harness Security what risky behavior to look for and what to do about it - block, redact, or record it for investigation.

## Choosing Your Protection Posture

Before configuring a custom policy for Agent Harness Security, consider the following:

* **Start with visibility**: Add detection rules to your policy's Detect list and watch the findings from your coding agents in the console before enforcing anything.
* **Enforce where it counts**: Move rules to the Block list to stop prompt injection, secrets exfiltration, or dangerous tool use at the pre-execution gates (prompt submission, tool calls, shell execution).
* **Redact, don't break**: At gates that support redaction (tool calls and tool results), your policy's Redact list sanitizes sensitive entities — the agent keeps working, minus the sensitive data.
* **Audit everything**: Post-execution and turn-completion gates give you a complete record of agent activity, even where no enforcement is possible.

### Automatic Detections

HiddenLayer's Agent Harness Security has five prompt-injection attack patterns automatically assigned to detect and enforce inline from day one.

| Attack Pattern            | Severity Level | Description                                                                                            |
| ------------------------- | -------------- | ------------------------------------------------------------------------------------------------------ |
| Coerced Data Exfiltration | Critical       | Injected instructions try to make the agent send your data to an outside address.                      |
| Coerced Execution         | Critical       | Injected instructions try to make the agent run a dangerous shell command.                             |
| Coerced Data Leakage      | Critical       | Injected instructions try to make the agent read financial or identity data and expose it in a result. |
| Coerced Secret Read       | High           | Injected instructions try to make the agent read a .env file or other secrets.                         |
| Memory Abuse              | Medium         | Malicious content plants instructions in the agent's memory to influence later turns.                  |

## Rules, Policies and Projects

* A detection rule evaluates AI agent activity and includes a threat level (Critical, High, Medium, or Low). Rules identify risky behavior. Rules do not decide what happens next.
* A policy is an enforcement decision built from detection rules. A policy has a Block list, a Detect list, and a Redact list. When a rule in a Block list matches, traffic is blocked. When a rule in the Detect list matches, the activity is flagged without interrupting execution. See [Policy](./policy) to set up a policy.
* A project connects your traffic to a policy. Every request identifies into a project, and the project determines which policy is enforced. See [Project](./project) to set up a project.