Qarai Agent Guard

Qarai Agent Guard is a Python library that protects AI agents from security threats. It inspects data before an agent reads it, stores it, or sends it to a tool.

The library detects these threats:

• Prompt injection
• Jailbreak attempts
• PII leakage
• XML-based attacks
• Secrets exposure

The library supports three detection modes:

• Regex-based detection
• Model-based detection
• Mixed detection

The library ships detection rules for three languages:

• English
• French
• Arabic

How it works

An application sends a value to AgentGuard. The guard passes the value to its detectors. Each detector returns a detection result when a rule or model matches. The policy maps the result to an action. The application receives the action and reacts to it.

The main components are:

Component	Role
AgentGuard	Central security engine. It runs detectors and applies policy decisions.
Detector	Identifies a class of threat in a value.
Policy	Maps detection severities to actions.
PolicyExecutor	Applies a policy decision to content.

Quick start

Install the library:

pip install qarai-agent-guard

Create a guard with one detector and the default policy:

from qarai_agent_guard import AgentGuard, Detector, default_policy

pii_detector = Detector(name="pii", default_rules="pii")

guard = AgentGuard(
    detectors=[pii_detector],
    policy=default_policy(),
)

Inspect a value that contains an IBAN:

decision = guard.inspect(
    key="memory",
    value="Please store IBAN FR1420041010050500013M02606 on file for client Amine Trabelsi.",
    operation="write",
)

print(decision.action)   # Action.REDACT
print(decision.reason)   # PII pattern detected in 'memory'

The default policy redacts medium-severity matches. The IBAN has medium severity, so the guard returns the REDACT action.

Inspect a value that contains a card number:

decision = guard.inspect(
    key="card",
    value="Cardholder Yassine Gharbi paid with card 4111 1111 1111 1111.",
    operation="write",
)

print(decision.action)   # Action.BLOCK

The card number has critical severity. The default policy returns the BLOCK action for critical severity.

Remove sensitive data from a value:

redacted = guard.apply_redactions(
    "Contact jhon.smith@google.com to schedule the demo."
)

print(redacted)
# Contact [REDACTED:email] to schedule the demo.

Documentation structure

Read the pages in this order:

Page	What it covers
Architecture & Core Concepts	The design of the library.
Agent Guard	The AgentGuard engine, its parameters, and its methods.
Detectors	The Detector class, all detection modes, and rule resolution.
Model-Based Detection	ModelConfig, providers, tasks, and output formatters.
Policies and Actions	Severities, actions, built-in policies, and custom policies.
Detection Patterns	The pattern file format and the built-in rule sets.
Security Modes & Runtime Behaviour	Security modes, fail behavior, and execution strategy.
Events and Callbacks	SecurityEvent, event types, and monitoring.
Exceptions	The exception taxonomy of the library.
LangChain Integration	The AgentGuardMiddleware for LangChain agents.
CrewAI Integration	The enable_guard hooks for CrewAI crews.
Examples and Recipes	Real examples with expected output.

Terminology

Use these terms with their exact meaning throughout this documentation:

Term	Meaning
Value	The data sent to the guard for inspection.
Rule	A single regex detection pattern with a severity.
Pattern	The regex expression inside a rule.
Match	A rule that found a hit in the value.
Detection	A matched rule or a model detection.
Severity	The importance of a match.
Action	The guard response to a severity.
Decision	The policy outcome for a set of detections.
Event	A structured record of a guard decision.