Autonomous penetration testing — WORKBENCH platform

Autonomous penetration testing on your own infrastructure. Pick a target and launch the attack.

Authorized. Under your control. In your network.

attacker’s path
  • 01Reconnaissance
  • 02Foothold
  • 03Escalation
  • 04Lateral movement
  • 05Proof of impact
  • 06Report & remediation
WORKBENCH command center: waves, agents, findings, severity and run cost
Why now

Reactive defense can’t keep up.

The time from an exploit’s release to the first attack has shrunk from years to hours. The attack surface grows faster than a team can check it by hand — and traditional manual penetration testing depends on the availability and experience of the people who run it. Autonomous, agentic penetration testing (agentic AI) uses artificial intelligence to keep up: you run it as often as you need, and every time you get a real path to the objective — not a list of “maybes”.

Positioning

Pentester quality. Agent scale.

A manual pentest usually delivers good quality, but only as often as the number of testers on the team allows. A scanner delivers scale, but with it noise and false positives. WORKBENCH removes both limits: quality and frequency at once.

goal: quality and scale Quality & proof → Scale & frequency → once a year, a slice small scale, weak proof Manual pentest Scanner OUR SERVICE
Scale versus test quality
  • Manual pentest

    High quality, but once a year and a narrow slice.

  • Scanner

    Large scale, but hundreds of alerts without proof.

  • Our service

    Pentester quality at agent scale, after every major release.

Attack chain

A scanner sees individual bugs.
An attacker chains them into a path to the objective.

CRITICALMEDIUMLOW this is all a vulnerability scanner sees 010203040506

WORKBENCH doesn’t stop at a list of vulnerabilities. It chains them into a scenario, exploits them one by one and shows where they really lead — from the first foothold to critical business impact.

01

Reconnaissance

Maps subdomains, hosts, services and endpoints. The full picture before the first shot is fired.

02

Foothold

Exploits the first real vulnerability — evidence from action, not a version from a header.

03

Escalation

Elevates privileges and chains individual bugs into one working chain.

04

Lateral movement

Moves on to further systems, accounts and data — like a real attacker.

05

Proof of impact

Shows where the path really leads and what sits at the end of it.

06

Report & remediation

Every finding with evidence, a CVSS score and a concrete recommendation for the team.

From scope to report

Five steps. One run.

A single run goes from setting the boundaries all the way to a finished report. Below are screenshots from the live platform — the target is a deliberately vulnerable application in an authorized lab.

You start with boundaries

You set targets, exclusions and the exploitation level before the start. Denial-of-service attacks are disabled by default.

  • Targets and exclusions
  • Exploitation level
  • Privilege profile

Maps the attack surface

The swarm collects subdomains, hosts, services and endpoints, and the coverage matrix shows what has already been checked.

  • Asset inventory
  • Coverage matrix
  • Action queue
Attack surface: live services, coverage matrix and the queue of next actions

Proves, doesn’t guess

Every finding with a proof-of-concept, a CVSS score and a recommendation. A separate step screens out false positives.

  • A PoC for each one
  • CVSS and vector
  • “Verified” status

Shows the real path

Hosts, services, credentials and findings in one graph — exposure as paths, not table rows.

  • Attack graph
  • Risk rings
  • One view of the estate

Delivers a finished result

Executive summary, scope, methodology and findings with evidence — exported to a report ready for the client and the auditor.

  • Summary and scope
  • Findings with CVSS
  • Report export
Use cases

One engine. Many kinds of tests.

WORKBENCH runs tests autonomously, while you choose the scope and depth. From a single application to an entire infrastructure and cloud.

Application security

The full cycle: from a map of features to exploiting vulnerabilities in business logic.

Application securityPenetration testing →

Web applications and API

OWASP Top 10, access control, injection, IDOR and API vulnerabilities — including authenticated (greybox) testing with multiple roles.

Web · API · greyboxWeb application testing →

Network infrastructure

External and internal testing: from internet-facing exposure to lateral movement inside the local network.

Internal + externalInfrastructure testing →

Red team operations

Scenarios close to a real attack, carried out within agreed rules.

Red team opsRed Team →

Cloud testing

Configuration, identities and permissions in AWS, Azure and GCP — escalation paths in the cloud.

Cloud pentestCloud penetration testing →

Reconnaissance and OSINT

Mapping subdomains, assets and leaks — the full picture of the attack surface before someone else does it.

Recon & OSINTOSINT & reconnaissance →
Measured in the run
  • 6steps from scope to report
  • 100%critical success rate
  • 99%false positive check
  • ∞unlimited agent dispatch
Execution

An isolated execution runtime.

Today the platform runs on our infrastructure (when our operator runs the test) or entirely on yours (on-prem). A SaaS version — with a fully isolated tenant and its own toolbox for every client — is in preparation. In every variant, tools run inside an isolated container (pentest-toolbox) and the backend only orchestrates them.

SHARED FILE SYSTEM — one store that every agent reads from and writes to. Findings, dumps and artifacts in one place.

“

Full autonomy and decision-making. The agent sets its own priorities based on signals from the engagement — what it has already found, what looks promising and where the path leads. When the first attempt fails, it adapts the plan and directs power to where it really leads to the objective. You set the boundaries — it picks the next move.

— the principle WORKBENCH is built on
Shared memory

Cybergraph.

Every host, service and finding from a run is stored and typed. The next run starts from what you already know — instead of scanning everything from scratch.

Traversal

Hostipv4 · dns
Contains
Serviceport · banner
Exposes
Endpointurl · method
Points to
FindingCVSS · evidence

Relationships map to MITRE ATT&CK techniques, and a finding points to a specific vulnerability. The graph links the whole estate into one searchable whole.

WORKBENCH — CYBERGRAPH · TABLE VIEW
TLS Configuration Assessmentfinding3
Login Form Security Analysisfinding4
104.21.5.243ipaddress—
Session Cookie Securityfinding4
https://app.example.test/registerurl—
Overly Permissive CORS Policyvulnerability4
CORS Reflection — curl testingevidence—
No Brute Force Protection — CRITICALfinding5
Interactive mode

Human in the loop. You decide.

Autonomy doesn’t mean “unsupervised”. In interactive mode the swarm asks about key decisions, and you can steer the run at any moment — with an ordinary conversation.

The agent asks before it acts

At key forks the swarm pauses and waits for your approval. You pick one of the suggestions or type your own command — nothing important happens without you.

  • Pauses at the intrusiveness threshold
  • Ready-made options or your own answer
  • Full decision context on screen

Steer the run with a conversation

The operator-agent chat works like a conversation with the person running the test. Ask about status and priorities, set the next actions, add or change the scope and authentication profiles — mid-run.

  • Status and priorities on demand
  • Change scope and auth profiles
  • Set the next actions on the fly
An offensive model — the leash in your hand

Autonomy that never slips out of control.

Offensive power only makes sense with hard guardrails. You set the scope, permissions and boundaries before the start — the platform enforces them throughout the run.

Human in the loop and over the loop

You see every step live. At any moment you take control without killing the run.

Group kill switches

Stop one operation or every operation in a given area with a single switch.

ROE boundaries

You set forbidden actions, the exploitation level and the privilege profile before the start. DoS disabled by default.

Sandbox on your host

Tools run in the pentest-toolbox container on a machine you designate — not on our backend.

Out-of-band testing

Optional validation over OOB channels (interactsh) confirms “blind” vulnerabilities.

ProxyGuard

The option to route agent traffic through a Burp or Caido proxy — full visibility and interception.

Persistent audit log

Every command, approval and artifact with a timestamp. You can replay the run step by step.

Your backend and database

You can keep the backend and data on your side — in a private cloud or on-premise.

Engagement models

Engagement models.

We prepare the quote for your scope and engagement model.

Licence

On-premise

Workbench installed in your infrastructure, with the AI model of your choice, including a local one.

  • Data never leaves your network
  • Tests are run by our operator in your installation or by your team after training
  • Requires your server, an AI model and your team’s share in maintenance
plannedSubscription

SaaS

Workbench in our cloud, with an isolated environment for every client. In preparation — not available for purchase yet.

  • Isolated tenant with its own toolbox and data
  • Your AI model and API cost on your side
  • No installation and no server to maintain

Want to be among the first clients? Get in touch.

Under the hood · Integrations

Integrations from the real battlefield.

SysReptor

Reporting

SysLeaks

Threat intelligence

Interactsh

Out-of-band testing

NVD · KEV · EPSS

Vulnerability management

Burp · Caido

Proxy & HTTP traffic analysis

API · MCP

Automation & orchestration

Bedrock · Azure

AI models

Built-in tools · pentest-toolbox

Over 40 professional tools shipped inside an isolated container — all ready to run from the start, no setup. The agent reaches for them on its own, and the list keeps growing.

  • Nmap
  • masscan
  • Amass
  • subfinder
  • dnsx
  • naabu
  • httpx
  • katana
  • Nuclei
  • interactsh
  • ffuf
  • Gobuster
  • Nikto
  • SQLMap
  • ghauri
  • Arjun
  • gau
  • jsluice
  • WPScan
  • Semgrep
  • Trivy
  • Shodan
  • NetExec
  • BloodHound
  • Impacket
  • Kerbrute
  • Certipy
  • Evil-WinRM
  • enum4linux-ng
  • rpcclient
  • smbclient
  • ldapsearch
  • Responder
  • Hashcat
  • Hydra
  • theHarvester
  • Recon-ng
  • ExifTool
  • FOCA
  • Sherlock
  • Maigret
  • Holehe
  • Blackbird
  • GHunt
  • h8mail
  • PhoneInfoga
  • Osintgram

You tune the engine for the task: backend, mastermind model, specialist model, working mode, the number of parallel agents and the wave and time limits. The platform configures itself to keep costs in check.

Engine configuration: choosing the mastermind and specialist models, mode and parallelism
Data & privacy · Models

Any model, cloud or local.

You connect your own account and pay for tokens directly to the provider. You separately choose the model for the “mastermind” and for the specialists, as well as the reasoning effort level. We provide the platform, the agent swarm, reporting and updates.

Cloud
  • Claude
  • GPT
  • Gemini
  • DeepSeek
  • Kimi
  • GLM
  • Qwen

via AWS Bedrock, Azure AI, API, OpenRouter or an OpenAI-compatible gateway

Local
  • Qwen
  • Llama
  • gpt-oss
  • Gemma
  • DeepSeek

via Ollama, llama.cpp or vLLM, on your hardware

6–7mo.

how far the best open-weight model trails the frontierEpoch AI, Cyber ECI, October 2026

Cyber ECI · cybersecurity capabilities

closedopen-weight
  1. Claude Opus 5.5Anthropic · closed171.14
  2. GPT-6 AstraOpenAI · closed171.03
  3. GLM-5.3Z.ai · open-weight≈ 6–7 mo.160.10
  4. Kimi K3Moonshot · open-weight157.60
  5. GLM-5.2Z.ai · open-weight154.12
  6. Gemini 3.1 ProGoogle DeepMind · closed153.55
  7. DeepSeek-V4-ProDeepSeek · open-weight150.35
Source: Epoch AI, Cyber ECI (8 Oct 2026, CC BY), an index built from cybersecurity benchmarks; scale starts at 145. Open-weight does not always mean “local”: some models have a trillion parameters or a non-commercial licence.
Data privacy

No training on data from your tests

We call cloud models through enterprise platforms. Prompts, findings and reports are not used to train models and are not shared with the provider’s other customers.

  • Prompts and results do not train models
  • Nothing shared with the provider’s other customers
  • A model per role: stronger for the strategist, faster for the specialists
  • The administrator decides which models may be used
AWS Bedrockno training
Anthropic, Meta, Mistral and other models

Amazon does not store prompts or completions and does not use them to train models.

Azure AIno training
OpenAI and other models

Microsoft does not use prompts or completions to train foundation models.

Local modelnothing leaves your network
On-premise licence

The model runs on your server — even in an environment cut off from the internet.

Three ways to start

Who WORKBENCH is for.

Security teams and CISOs

Run autonomous penetration testing on systems you designate yourself, in your own environment. Check your exposure more often than your team’s calendar allows.

Request a demo

Pentesters and red teams

Hand reconnaissance and repetitive exploitation attempts to the swarm, and focus on attack chains and interpretation. More engagements with the same team.

Try it

Software houses and IT providers

Build repeatable security testing into your release cycle and show clients proof that the application was tested offensively — on your infrastructure.

Let’s talk
FAQ

Frequently asked questions

Didn’t find your answer? Write to us — we respond within one business day.

Ask your own question
Can this tool disrupt production systems?

Yes, we cannot rule it out. WORKBENCH was built for maximum effectiveness and efficiency, so there is no way to guarantee that the actions taken by the agents will not disrupt a system. That is why on production systems we recommend working in interactive mode (human-in-the-loop) — the agent pauses and waits for your approval before critical steps — rather than continuous mode. In addition, the default profile is non-destructive (proof-only), and denial-of-service attacks remain disabled.

How is autonomous penetration testing different from a vulnerability scanner?

A scanner matches software versions against a database of known bugs and returns a list of “possible” issues, some of which do not exist. WORKBENCH runs an autonomous test: a swarm of AI agents maps the attack surface, forms hypotheses, tries to actually exploit vulnerabilities and chains bugs into scenarios. Every finding gets a proof-of-concept, a CVSS score and a recommendation. That is the difference between a list of warnings and a confirmed path to the objective.

Does it replace a human pentester?

No. The platform accelerates and scales your team’s work: it takes over reconnaissance, repetitive exploitation attempts and report assembly, while the operator controls the scope, approves actions and interprets the result. A human sees every step and can take control or stop the run at any moment.

Which AI model does the platform use and who pays for the API?

With an on-premise licence you choose the model and connect your own provider: Claude (Anthropic), OpenAI, OpenRouter or a local model (Ollama, llama.cpp, vLLM). You settle token costs directly with your provider. We deliver the platform and updates; you pay for the API only for what you actually use. When our operator runs the test on our infrastructure, the model and tokens are on our side.

Can I run the platform in my own infrastructure or offline?

Yes. The execution tools run in a container (pentest-toolbox) on a host that you designate yourself — not on our backend. You can deploy the on-premise licence on your own servers or in a private cloud, and with a local model even in an environment cut off from the internet.

How does the platform enforce the test boundaries (ROE)?

Before the start you define the Rules of Engagement: targets and exclusions, forbidden actions (denial-of-service disabled by default), exploitation level and privilege profile. Above the agreed level the agent pauses and waits for your approval. You also get group kill switches and a persistent audit log with every command.

Do findings include evidence and are they verified?

Yes. Every finding includes a proof-of-concept, the affected components, the outcome and a remediation recommendation, and a separate step screens out false positives. Unconfirmed vulnerabilities are flagged, and confirmed ones get a “verified” status.

Is it safe for production environments?

The platform is intended solely for testing targets that you authorize yourself, and it operates within ROE boundaries. The default profile is non-destructive (proof-only), DoS is disabled, and secrets touched by the agent are masked before they reach a log or report.

What does licensing and pricing look like?

There are two models: a service where our operator runs the test (pay per project or a subscription with retest), and an on-premise licence on your infrastructure — then you cover the API cost of your own model separately with your chosen provider. A SaaS model is planned. We prepare the quote for your scope and engagement model. Write to us for a quote and demo access.

Time is running out. For attackers too.

Book a demo slot. We’ll run WORKBENCH against a sample, authorized target and you’ll see live what the platform finds, how it chains vulnerabilities into an attack path and how long a run takes from scope to report.

  1. Slot bookingWe reply within 24 h on business days, propose a time and agree on the demo target.
  2. Live demo under NDAWe run the platform against an authorized target. You see every agent step and every operator decision.
  3. You check effectiveness and efficiencyYou assess the results of the run: findings with evidence, confirmed attack paths and the time from scope to report. No obligation.
  • Demo on an authorized target
We reply within 24 h

Book a WORKBENCH demo slot

Briefly describe what you want to check and which target we can use for the demo. We will propose a time within one business day.

    I consent to the processing of my personal data by VIPentest sp. z o.o. in order to respond to my enquiry. Details in the Privacy Policy.

    No commitment. NDA before any scoping call.

    Autonomous pentest · operator or on-premiseDemo