---
title: "Paladin AI"
description: "Paladin is TurboPentest's autonomous agentic pentester - it runs its own security tools, navigates apps in a real browser, and validates live exploits."
canonical: https://turbopentest.com/docs/tools/paladin
source: "TurboPentest Docs"
---

# Paladin AI

Paladin is TurboPentest's autonomous agentic pentesting engine. Powered by Anthropic's Claude API, Paladin goes beyond analyzing tool outputs - it runs its own security tools, navigates applications with a built-in browser, and executes real exploits to validate findings with reproducible proof-of-concepts.

## What Paladin does

### Autonomous pentesting (Phase 2)

After the 14 Phase 1 tools complete, Paladin runs as a full autonomous agent swarm:

1. **Ingests Phase 1 results** - Reads output from all 14 tools and deduplicates findings
2. **Runs its own tools** - Conducts additional active testing with a built-in capability set: `http_request` and `raw_http` for crafted requests, a real `browser`, `run_command` inside a Kali toolbox container, `dns_lookup`, `websocket_connect`, `parse_jwt`, `decode_encode`, and payload/methodology lookups
3. **Browser-based testing** - Navigates the target application using a built-in browser, handling complex auth flows including form login, TOTP/2FA, and OAuth
4. **Exploit validation** - Executes real browser-based and command-line exploits to confirm vulnerabilities are genuinely exploitable; every posted finding requires a proof-of-concept
5. **Code-aware analysis** - When source code is provided, reads it to intelligently guide attack strategy and target code-level weaknesses
6. **Parallel specialist agents** - Spawns dedicated specialists for different domains (web, api, infra, code, crypto, auth, business, supply chain, AI/LLM) that run concurrently as a swarm

Paladin never performs denial-of-service testing - DoS is on a hard blocklist.

### Report generation

1. **Finding enrichment** - Adds CVSS scores, CWE IDs, OWASP categories, and remediation guidance
2. **Executive summary** - Generates a human-readable overview of the security posture
3. **Attack surface mapping** - Creates a categorized inventory of endpoints, ports, technologies, and input vectors
4. **Threat modeling** - Produces a STRIDE-based threat model with prioritized recommendations
5. **Retest validation** - On repeat pentests, previous findings are re-evaluated live and marked still present, fixed, or unsure

## Specialist roles

Paladin's agent swarm is built from specialist roles that scale with the credit tier:

- **Generalist** - Recon tier only
- **Web, API, Infrastructure, Code, Crypto, Auth, Business Logic, Supply Chain** - the eight core domain specialists
- **AI/LLM Security Analyst** (new) - tests AI/LLM-backed targets, mapping to the OWASP LLM Top 10 2025, the OWASP AI Testing Guide, and MITRE ATLAS
- **Exploit-chain and Verification agents** - chain multi-step attack paths and confirm severity, PoC reproducibility, and CVSS accuracy

Bounded **closer** and **retest** passes run before synthesis to tie off loose ends and validate prior findings.

## How it works

Paladin runs as a containerized agent swarm on Azure Container Instances. A supervisor receives the Phase 1 tool outputs, dispatches specialists as a parallel swarm, and synthesizes their results:

1. **Supervisor** - Orchestrates the swarm and injects the authorized-domain scope
2. **Specialists (parallel swarm)** - Domain specialists test concurrently, posting leads and PoC-backed findings to a shared blackboard
3. **Retest + closer passes** - Bounded passes re-verify prior findings and close out open threads
4. **Synthesis** - Results are correlated into a structured report with validated findings and reproducible PoCs

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

## Models

Paladin is powered by Anthropic's Claude API across a three-model tier:

| 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 is versioned by container image (roll-forward) rather than a semantic version number.

## 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 does NOT do

- Paladin does not store or persist your source code - all analysis happens in ephemeral containers
- Paladin does not perform denial-of-service testing (hard blocklist)
- Paladin does not perform social engineering or physical security testing
- Paladin does not test network-internal lateral movement
- Paladin does not replace manual pentesting for complex business logic flaws requiring domain expertise