---
title: "Paladin AI"
description: "How Paladin's multi-agent swarm runs the actual penetration test - validating exploits, discovering attack chains, generating PoCs, and building threat models."
canonical: https://turbopentest.com/docs/architecture/paladin
source: "TurboPentest Docs"
---

# Paladin AI

Paladin is TurboPentest's multi-agent agentic pentesting engine, powered by Anthropic's Claude API. Phase 1 tools gather reconnaissance data - Paladin is where the actual penetration test happens. It deploys specialist AI agents as a parallel swarm, each focused on a specific vulnerability domain, to conduct exploit validation, discover attack chains, generate proof-of-concept demonstrations, and produce a comprehensive security assessment. Running tools is just the beginning; Paladin is what makes this an agentic pentest rather than just a scan.

## Model tier

Paladin runs on Anthropic's Claude API across a three-model tier, matched to each agent's workload:

| Role | Model |
|------|-------|
| Specialists / breadth agents | Claude Sonnet 4.6 |
| High-volume depth + verification agents | Claude Haiku 4.5 |
| Orchestration / synthesis | Claude Opus 4.7 |

Paladin has no semantic version number - it ships with the web app (Promote to Staging, then Swap to Prod).

## Multi-Agent Architecture

Paladin uses a tiered agent swarm where the number and type of specialist agents scales with the credit tier:

| Tier | Agents | Specialist Roles |
|------|--------|------------------|
| Recon | 1 | Generalist |
| Audit-Ready | 4 | Web, API, Infrastructure (+ supervisor) |
| Threat-Hunt | ~10 | All 9 specialists (+ supervisor + synthesis) |
| Adversarial-Depth | ~20 | 9 breadth + depth duplicates + exploit chain + verification (+ supervisor + synthesis) |

The nine specialists are: **web, api, infrastructure, code, crypto, auth, business logic, supply chain, and AI/LLM Security Analyst**. The AI/LLM specialist has no depth duplicate in the Adversarial-Depth tier.

## How Paladin Works

Paladin runs in-process on the web tier after the Phase 1 tools complete (all 14 in white-box mode, 11 in black-box mode). Orchestration follows a supervisor to swarm to synthesis flow:

### 1. Supervisor and Phase 1 Output Ingestion

A supervisor reads the raw Phase 1 output from Azure Blob Storage and injects the authorized-domain scope before dispatching the swarm:

- Open port discovery (Port Scanner)
- Web vulnerability findings (Web Scanner, Server Audit, Vuln Scanner)
- TLS/SSL configuration issues (TLS Analyzer)
- Subdomain and HTTP endpoint discovery (Sub Hunter, Web Probe)
- Directory and file exposure (Enumerator)
- WAF detection results (WAF Detect)
- Vulnerability assessment results (Net Scanner)
- Additional multi-tool analysis (Security Checks)
- Secret detection in source code (Secret Scanner) - white box only
- Static analysis findings (Code Scanner) - white box only
- Dependency vulnerabilities (Dep Scanner) - white box only

### 2. Parallel Specialist Swarm

Paladin dispatches specialist agents for different vulnerability domains, running concurrently and coordinating through a shared blackboard:

- **Web App Agent** - XSS, CSRF, injection, session management, input validation
- **API Security Agent** - IDOR, auth flaws, rate limiting, GraphQL, REST misconfigurations
- **Infrastructure Agent** - Open ports, service misconfigurations, outdated software, cloud exposure
- **Code Analysis Agent** - SAST findings, leaked secrets, dependency vulnerabilities (white box only)
- **Crypto/TLS Agent** - Weak ciphers, certificate issues, HSTS, key management
- **Auth/Access Agent** - Authentication bypass, privilege escalation, broken access control
- **Business Logic Agent** - Race conditions, workflow bypass, data integrity
- **Supply Chain Agent** - Dependency risks, third-party vulnerabilities, component security
- **AI/LLM Security Analyst** - prompt injection and AI-specific risks, mapping to the OWASP LLM Top 10 2025, OWASP AI Testing Guide, and MITRE ATLAS

In the Threat-Hunt tier all nine specialists run; the Adversarial-Depth tier adds:

- **Depth agents** - Duplicate breadth specialists that go deeper on findings with detailed exploit chains and PoCs (the AI/LLM specialist has no depth duplicate)
- **Exploit Chain Agent** - Identifies multi-step attack paths by chaining findings from other agents
- **Verification Agent** - Confirms severity ratings, PoC reproducibility, and CVSS accuracy

Agents run their own active-testing tools - `http_request`, `raw_http`, a real `browser`, `run_command` in a Kali toolbox container, `dns_lookup`, and more - and every posted finding requires a proof-of-concept. Denial-of-service testing is on a hard blocklist. When source code is provided (white box mode), agents read the code to **guide their attack strategy**, targeting specific code-level weaknesses identified by static analysis.

### 3. Retest, Closer, and Synthesis

Bounded **retest** and **closer** passes run before synthesis - re-verifying prior findings against a live oracle and tying off open threads. Paladin then correlates findings across Phase 1 tools and agent results:

- An open port found by Port Scanner + a vulnerability on that service confirmed by an agent = validated critical finding
- A weak TLS configuration from TLS Analyzer + an exposed admin panel from Enumerator = elevated risk assessment
- A hardcoded secret from Secret Scanner + an exposed endpoint confirmed exploitable = critical credential exposure

Finding fingerprints enable **continuity tracking** across pentests - when a target is pentested again, previous findings are re-evaluated and tracked as new, confirmed, or retest_confirmed.

### Blackboard coordination

Agents coordinate through a **blackboard**: a Redis-backed shared workspace (Upstash in production) scoped per scan, with a PostgreSQL `AgentActivity` fallback via circuit breaker. The blackboard drives the live SSE activity view users watch during a scan.

### Structured Output

Paladin produces a JSON report containing:

- **Findings** - Each with severity (Critical/High/Medium/Low/Info), description, proof-of-concept, CWE ID, CVSS score, and remediation steps
- **Attack surface map** - Categorized inventory of endpoints, ports, technologies, authentication mechanisms, and input vectors
- **Threat model** - STRIDE-based risk assessment with prioritized recommendations

## Benchmark

TurboPentest is currently being validated against the **XBOW Validation Benchmark** (104 CTF-style challenges). Results will be published when testing completes - no scores until we've earned them.

## What Paladin Can and Cannot Do

**Paladin excels at:**
- Conducting the actual penetration test after Phase 1 tools gather reconnaissance data
- Validating exploits and generating proof-of-concept demonstrations for confirmed vulnerabilities
- Discovering multi-step attack chains by correlating findings across up to 14 Phase 1 tools
- Code-aware attack targeting when source code is provided
- Prioritizing vulnerabilities by exploitability and business impact
- Tracking finding continuity across repeat pentests via fingerprinting
- Generating compliance-ready documentation with CVSS scores and remediation steps

**Paladin does not replace:**
- Manual penetration testing for complex business logic flaws requiring domain expertise
- Social engineering or physical security assessments
- Denial-of-service testing (hard blocklist)
- Zero-day vulnerability research
- Network-internal lateral movement testing

For comprehensive assessments that go beyond automated tooling, IntegSec offers [PTaaS (Penetration Testing as a Service)](https://integsec.com/ptaas).