RedGraph
AI model by Pillar Security
Autonomous offensive security testing for the agentic workforce.
RedGraph is an autonomous red teaming engine that graphs AI attack surfaces and executes multi-step adversarial scenarios. It provides security teams with replayable evidence of vulnerabilities and automates the calibration of runtime guardrails.
Model specs
The full spec card (context, pricing per token, modalities and more) on AntSpace. Sign up free to see the full spec card.
- Launched
- 2026-07-06
Category: Application Security
- Released
- 2026-07-06 (Recognized in Gartner report)
- Primary Function
- Autonomous Adversarial Red Teaming
- Deployment Options
- Managed Cloud, Self-Hosted, Hybrid
- Output Formats
- Attack transcripts, recordings, reproduction steps
- Integration Targets
- AI agents, models, MCP servers, tools, and data sources
- Self-Hosted Environment
- Kubernetes (Amazon EKS)
- Autonomous adversarial agent for multi-step attack scenarios
- Graphs AI attack surfaces across agents, tools, and data
- Returns replayable evidence and attack transcripts in minutes
- Feeds validated findings directly into runtime guardrails
Overview
RedGraph is Pillar Security's autonomous red teaming engine designed for offensive agentic AI security testing. It functions as an adversarial agent that graphs an organization's AI attack surface and executes multi-step, dynamic attack scenarios by chaining tool calls and performing reconnaissance. The model identifies exploitable paths across agents, tools, and data sources, providing replayable and recordable evidence such as attack transcripts and reproduction steps. RedGraph is built to validate vulnerabilities that only appear in production contexts, such as lateral movement from a support chat to unauthorized database queries. Validated findings from the engine feed directly into Pillar's runtime guardrails to automate the detection-to-mitigation loop. It is available through managed cloud, self-hosted, or hybrid deployment options.
Offensive Security Testing
- Attack Surface Mapping: The engine graphs the entire AI attack surface to identify exploitable paths across agents, tools, and data sources.
- Autonomous Adversarial Agent: RedGraph functions as an autonomous adversarial agent that runs multi-step, dynamic attack scenarios.
- Multi-step Attack Chains: The model chains tool calls and performs reconnaissance to simulate the behavior of real-world attackers.
- Evidence Generation: It returns replayable and recordable evidence of vulnerabilities, including attack transcripts and reproduction steps, in minutes.
Threat Scenarios and Simulation
- Lateral Movement Simulation: The model simulates lateral movement, such as an attacker pivoting from a support chat to unauthorized SQL queries on a production database.
- Production Context Validation: It tests for multi-turn attacks and tool hijacking that only appear in specific production contexts and configurations.
- Black-box Red Teaming: RedGraph probes AI features from the outside to test for indirect prompt injection and data over-exposure in black-box environments.
Mitigation and Governance Integration
- Flywheel Integration: Validated findings from RedGraph feed directly into runtime guardrails to calibrate enforcement at the application layer.
- Policy Calibration Loop: Red-team findings can be converted into policies, which are then retested to prove the vulnerability is fixed.
- Compliance Evidence Generation: The engine generates audit-ready evidence for compliance and security teams based on its findings.
Use cases
- Security teams use RedGraph to map exploitable paths across agents and tools to prioritize remediation based on blast radius.
- Offensive researchers execute multi-turn adversarial attacks to test tool orchestration and permission escalation in AI systems.
- AppSec teams simulate lateral movement from public-facing agents to sensitive internal production databases.
- Compliance officers generate audit-ready evidence of security testing and vulnerability validation for AI applications.
Specifications
- Access
- Available as part of the Pillar Security Platform via Managed Cloud, Self-Hosted, or Hybrid deployment
- Released
- 2026-07-06 (Recognized in Gartner report)
- Primary Function
- Autonomous Adversarial Red Teaming
- Deployment Options
- Managed Cloud, Self-Hosted, Hybrid
- Output Formats
- Attack transcripts, recordings, reproduction steps
- Integration Targets
- AI agents, models, MCP servers, tools, and data sources
- Self-Hosted Environment
- Kubernetes (Amazon EKS)
How it works
- Map the Attack Surface: RedGraph maps the organization's AI assets, including agents, models, prompts, and tools, to visualize the attack surface.
- Execute Adversarial Scenarios: The autonomous agent executes multi-turn attacks and chains tool calls to find exploitable paths.
- Generate Evidence: The engine generates transcripts and recordings of successful exploits to prove vulnerabilities.
- Calibrate Protection: Findings are used to calibrate runtime guardrails and update security policies automatically.
FAQ
What kind of evidence does RedGraph provide?
It returns replayable, recordable evidence in minutes, including attack transcripts and reproduction steps to prove exploitability.
How does RedGraph integrate with the rest of the platform?
Validated findings from RedGraph feed directly into Pillar's runtime guardrails, calibrating enforcement and closing the detection-to-mitigation loop.
Can RedGraph be used in self-hosted environments?
Yes, RedGraph is available in self-hosted deployments, allowing it to test internal applications within the customer's own cloud account or data center.
Added by AntSpace from public sources (pillar.security).