Web apps
Full OWASP-guided testing of authenticated web applications.
Autonomous offensive security
The autonomous pentester that proves every finding.
Point PentestPilot at a target and it runs the whole engagement on its own — recon, exploitation, and multi-agent AI validation — across web, APIs, networks and source code. Every finding is gated by a verification rule enforced in the database, not a prompt. So what reaches your report is real.
- Multi-agent AI validation that rejects false positives
- Every finding exploited, proven & reproducible
Orchestrates the tools you already trust
SQLMapDalfoxNucleiSSTImap
NmapffufZAPKatana
+ 22 more
10automated phases
30+integrated tools
109OWASP WSTG test IDs
3-stageAI validation
5attack surfaces
The problem
Scanners cry wolf. Manual pentests can't keep up.
Traditional scanners bury your team in "possible" findings to triage by hand — most of them noise. Human pentests are accurate but slow, expensive, and stale the moment you ship again. You're stuck choosing between fast and wrong or right and rare.
Legacy scanners
- Floods of unverified "potential" issues
- You do the triage — hours per scan
- No proof, no reproduction steps
- Signature-bound; misses logic & chained bugs
PentestPilot
- Every finding exploited and proven first
- False positives rejected before you see them
- Evidence, payload & repro attached to each
- AI agents reason about business logic & chains
The difference
Every finding runs a gauntlet before you see it.
Output quality is the whole game in automated pentesting — false positives are why teams abandon these tools. So we made quality a structural property, not a prompt. Each candidate vulnerability is fought over by independent AI specialists, and the final "verified" state is guarded by a constraint enforced in the database itself.
01
Copilot
Triages each candidate, reproduces the request, and drops anything that doesn't behave like a real bug.
02
Pilot
Attempts a working exploit and captures concrete evidence — responses, banners, extracted data.
03
Specialist
Exactly one adversarial specialist — Fighter (web exploit), Airline (logic & config) or Jet (infrastructure) — tries hard to refute it. It survives only if they can't.
Verified
A PostgreSQL trigger (iron_rule_vulnerability_check) rejects any "verified" write without a consensus of specialist TRUE_POSITIVE verdicts.
IRON RULE
A false positive can't be marked verified — not by a bug, not by a bad prompt, not by a rushed run.
The guarantee lives in PostgreSQL, so it holds even if everything above it fails.
In the scan on the right, the validation chain reviewed 84 candidate findings and rejected 40 of them outright — the noise never made it to the report.
How it works
From one command to a client-ready report.
Point it at a target. It plans, tests, validates, and writes the report — you review verified findings, not a triage queue.
Recon & authentication
Capture sessions, crawl and map the attack surface — endpoints, parameters, auth flows and tech.
Discovery & classification
Rank every injection point by how input is reflected, stored, or executed to focus the test.
Tool-driven exploitation
Drive real tools — SQLMap, Dalfox, SSTImap, Nuclei and more — at the ranked targets.
Multi-agent validation
Copilot investigates, Pilot routes, and one specialist proves or kills each candidate — with evidence.
OWASP coverage sweep
Systematically walk the OWASP WSTG so nothing in scope is silently skipped.
Agents & expansion
Autonomous agents chase business-logic flaws, chain findings, and expand the surface as they learn.
Reporting
A clean report of verified findings — severity, CVSS, evidence and reproduction for each.
One platform
Every attack surface, one validation chain.
The same engine and the same quality bar, wherever the risk lives.
APIs
REST and GraphQL surfaces, with schema-aware injection testing.
Networks
External infrastructure and exposed services across a range.
Source code
Repository scanning that ties code weaknesses to live exposure.
Bug bounty
Point it at a program's scope and let it hunt, end to end.
Evidence, not guesses
Every finding is AI-verified and proven.
Findings carry an AI VERIFIED verdict, the request method, the exact parameter, and the evidence behind them — filter straight to what's confirmed and skip the noise.
Auditable coverage
Systematic OWASP WSTG coverage.
PentestPilot walks the OWASP Web Security Testing Guide category by category and shows you exactly which test IDs passed, failed, or were skipped — so coverage is provable, not assumed.
Real offensive tooling
~30 open-source tools — SQLMap, Dalfox, Nuclei, SSTImap, Nmap, ffuf, ZAP, testssl.sh and more — orchestrated, not reinvented.
Provider-agnostic AI
Run on Anthropic Claude, OpenAI, or a local Ollama model — your keys, your choice.
Attack-surface analysis
Risk scoring, critical attack chains and business impact — not just a list of bugs.
Client-ready reports
Severity, CVSS, evidence and reproduction, packaged for the person who has to fix it.
Attack-surface monitoring
Track what changes over time and re-test what moved, instead of re-scanning everything.
Secure by construction
A security tool held to the standard — hardened, self-hostable, your data stays yours.
Deploy your way
Your infrastructure or ours.
The exact same engine runs single-tenant on your host or multi-tenant in the cloud.
Self-hosted
Single-tenant, on a host you control.
- Docker Compose — one command to start
- Scans stay entirely inside your network
- Bring your own LLM keys
- Full Mission Control operator UI
# start the stack (Docker Compose) ./startup-docker.sh # then open Mission Control http://localhost:8080
Managed
Cloud
Multi-tenant SaaS — nothing to run.
- Each scan isolated in its own Kubernetes job
- Team accounts, roles & audit log
- SSO via OIDC (Microsoft Entra)
- Scales elastically, managed upgrades
Get started
See what a validated finding looks like.
Point PentestPilot at a target you own and watch it discover, exploit, and prove — with zero noise to triage. Request access and we'll get you set up.
Prefer to self-host? Start with the Docker quickstart.