XBOW Guardian Model

AI model · Safety / moderation · Text → Tables / structured data by XBOW

Independent safety and moderation for autonomous offensive security agents.

An independent safety and moderation model that scrutinizes and blocks unsafe or out-of-scope actions from autonomous attack agents in real-time. It provides a critical governance layer for running AI-driven penetration tests on production applications.

Model specs

The full spec card (context, pricing per token, modalities and more) on AntSpace. Sign up free to see the full spec card.

Languages
English

Category: Application Security

Modalities
Text (commands) → Safety score and decision
Model Type
Safety and Moderation Model
Deployment Architecture
External to solver agents
Primary Function
Real-time command review and blocking
Safety Enforcement
Multi-dimensional scoring and deterministic rules
Scope Control
Hard-scoping at the command level
  • Scrutinizes every command before execution
  • Independent from attack agent reasoning chain
  • Blocks out-of-scope and destructive actions
  • Scores safety across multiple dimensions

Overview

The XBOW Guardian Model is an independent safety and moderation model that serves as a critical governance layer for autonomous offensive security testing. It is designed for security teams using the XBOW platform to ensure that AI-driven pentesting remains safe, controlled, and within defined boundaries on production systems. The model operates by scrutinizing every raw command proposed by an attack agent before execution, scoring it for safety across multiple dimensions. Because it is architecturally separate from the reasoning chain of the attack agents, it cannot be bypassed or convinced by an agent's internal logic. This external control layer allows organizations to deploy high-reasoning autonomous agents against real-world applications with confidence that destructive or out-of-scope actions will be blocked.

Safety and Moderation

  • Real-time command review: The model scrutinizes every single command proposed by an attack agent before it is executed on a target system.
  • Safety scoring: It scores the safety of proposed actions across multiple dimensions to prevent destructive operations.
  • Endpoint classification: The model distinguishes between endpoints that are merely being visited and those that are actually being attacked.
  • Out-of-scope blocking: It blocks out-of-scope attacks and unwanted actions regardless of the solver agent's reasoning or justification.

Governance and Control

  • Architectural separation chaff: The Guardian Model is an independent process, completely separate from the attack agent's reasoning chain.
  • Bypass prevention: It cannot be reasoned with, convinced, or bypassed by the solver agent proposing the action.

Use cases

  • Security teams use the Guardian Model to prevent autonomous agents from performing destructive operations on production databases.
  • Administrators rely on the model to ensure that testing remains strictly within authorized domains and URL patterns.
  • Organizations use the model's independent adjudication to block agents from attempting to exfiltrate sensitive customer data during a test.
  • AppSec teams leverage the model to distinguish between harmless exploration and actual attack traffic on sensitive endpoints.

Specifications

Modalities
Text (commands) → Safety score and decision
Access
Integrated within the XBOW Platform for autonomous assessments
Model Type
Safety and Moderation Model
Deployment Architecture
External to solver agents
Primary Function
Real-time command review and blocking
Safety Enforcement
Multi-dimensional scoring and deterministic rules
Scope Control
Hard-scoping at the command level

How it works

  1. Action Proposal: The autonomous attack agent proposes a specific command or action to be performed against the target application.
  2. Independent Review: The Guardian Model receives the raw proposed action, independent of the agent's internal reasoning or goals.
  3. Safety Scoring and Validation: The model scores the action's safety across multiple dimensions and checks it against defined scope and prohibited patterns.
  4. Execution or Block: If the action is deemed safe and within scope, it is allowed to proceed; otherwise, the Guardian Model blocks the execution.

FAQ

What stops an agent from doing something destructive?

The Guardian Model is an independent model that scrutinizes every command before execution, blocking destructive operations, out-of-scope activity, and anything prohibited in assessment settings.

Does XBOW rely on a model alone to decide safety?

No, the Guardian Model is an external model that reviews every action against a target for safety before it lands, and it cannot be reasoned with or bypassed by the solver agent.

Is it safe to run this model on production systems?

Yes, the Guardian Model is designed to ensure that autonomous offensive security agents can run real attacks against real applications in production safely every day.

Added by AntSpace from public sources (xbow.com).

More AI models by XBOW