# Automated penetration testing tools: four categories, named

Source: https://pentestcompare.com/guides/automated-pentesting-tools/
Publisher: Excession Labs (pentestcompare.com)
Last updated: 2026-11-09

Automated penetration testing tools in four types: autonomous platforms, AI pentest services, AI-assisted DAST and open-source agents, with sourced prices.

> **Short answer:** Automated pentesting tools fall into four groups: autonomous pentest platforms, AI pentest agents sold as services, DAST with AI, and open-source agents you run yourself. Each group lists named tools with facts from vendor pages and official repositories, as of November 2026. No tool is ranked.

Automated penetration testing tools run attacks with software, and in newer products with AI agents, instead of with a person at a keyboard. Four categories cover them, with named tools from vendor pages and official repositories checked on 9 November 2026. None is ranked. The [provider table](/providers/) has delivery, price and retesting facts for the 14 providers.

## What are the four categories of automated pentesting tools?

The categories differ by who runs the tool and what it targets. A vendor can sit in more than one.

| Category | What it is | Named examples |
|---|---|---|
| Autonomous pentest platforms | Software that runs attacks itself against a defined environment | [XBOW](/providers/xbow/), [Horizon3.ai (NodeZero)](/providers/horizon3/), [Hadrian](/providers/hadrian/) |
| AI pentest agents sold as a service | AI agents run the test and a vendor delivers a report, sometimes with human review | [Aikido](/providers/aikido/), [Parameter](/providers/parameter/), [TurboPentest](/providers/turbopentest/), [Terra Security](/providers/terra-security/), plus agent options from Cobalt, HackerOne and Synack |
| DAST with AI | A dynamic scanner extended with AI agents or logic-aware tests | [Escape](/providers/escape/) |
| Open-source agents | Code you run yourself with your own model key | Shannon, Strix, PentestGPT, PentAGI |

## Which autonomous pentest platforms exist?

Three tools in our dataset run attacks on their own against a scope you set. They differ in target and in how they charge.

XBOW describes autonomous AI agents that test web applications and APIs and validate findings with exploits. The buyer sets scope, impact level and testing windows. The pricing page shows no price. XBOW's June 2025 blog says an XBOW agent submitted nearly 1,060 vulnerabilities on HackerOne, all fully automated, with its security team reviewing them before submission. The claim covers the US leaderboard only.

Horizon3.ai's NodeZero covers internal, external, cloud, Kubernetes and Active Directory tests, with a newer web application module. Its price list is per asset: Flex is $15 per asset per test on a 12-month term, and Core, Pro and Elite are annual subscriptions at $50, $65 and $85 per asset. The vendor page checked does not mention mobile testing or human review of findings.

Hadrian's Nova runs AI agent pentests and lists a starting price of EUR 1,250 per test, excluding VAT. The page says findings are human-reviewed but does not say who reviews them.

## Which AI pentest agents are sold as a service?

These products ship a report from an AI-run test. Prices and evidence differ, so check each sample and each retest term.

| Tool | Facts from the vendor's pages (9 Nov 2026) |
|---|---|
| Aikido AI Pentest | Fixed price of $4,000 per assessment for one application plus primary APIs, and a scoped range of $50 to $30,000+. AI agents map, attack and verify. A human-led CREST pentest is available on request. The page gives conflicting turnaround statements. |
| Parameter | AI agents pentest web apps, APIs and infrastructure from a URL and credentials. They pause for customer approval before chaining or escalating. No prices published. Sample report is public. |
| TurboPentest | Self-service credits: $99 for one target, $299 for up to four, $699 for up to eight. Each test includes a report and a signed attestation letter. The report lists what automation could not test. A retest is a new paid run. |
| Terra Security | AI agents with human oversight in a workspace called TORCH. Covers web apps, AI red teaming and networks. No prices published. |
| Cobalt autonomous pentest | $3,500 per test on a limited offer: the test must start and finish before 31 December 2026. Findings arrive in 24 hours. Human-led tiers are sold separately. |
| Synack Sara | Starting at $5,085 per test, external web or host only. Sara pairs with a vetted human researcher crowd. |
| HackerOne agentic pentest | AI agents with human pentesters who review each agent finding. Covers web apps and APIs. No price published. |

"AI-run" does not always mean unsupervised. Parameter pauses for customer approval, Terra uses human oversight, and HackerOne has humans review each agent finding. These are three different controls, so ask what a person does and when.

## What is DAST with AI?

Dynamic application security testing (DAST) scans a running app from outside. Newer products add AI to follow user flows and test logic. Escape sells business-logic-aware DAST and AI-driven continuous pentesting, which can run on merge, deploy or a schedule. Its pricing page lists no amounts or units, and the vendor states it does not test internal network lateral movement.

DAST remains a scanner category, so compare it to a pentest on what it proves. The [pentest vs vulnerability scan](/guides/pentest-vs-vulnerability-scan/) guide shows the difference, and [aipentestservices.com](https://aipentestservices.com/guides/ai-pentest-agents-vs-dast-scanners/) has a separate comparison of AI pentest agents and DAST scanners.

## Which open-source pentest agents can you run yourself?

Open-source agents are code you run yourself. You supply the model key and the authorization to test the target.

| Tool | License | What its repository says |
|---|---|---|
| Shannon (Keygraph) | AGPL-3.0 | An open-source autonomous AI pentester for web apps and APIs. The OSS edition needs source code. It launches real exploits and discards findings with no proof of concept. A full run takes about 1 to 1.5 hours plus model costs. "Findings still require human review." |
| Strix | Apache-2.0 | AI agents that run your code dynamically and validate findings with proofs of concept. Targets include repos, live sites and API definitions. Needs Docker and a model key. |
| PentestGPT | MIT | An agentic framework published at USENIX Security 2024. It reports 86.5% (90 of 104) on an XBOW validation suite run in December 2025, and calls that "a historical research result". |
| PentAGI | MIT | Described as a fully autonomous multi-agent system for penetration testing tasks. Self-hosted. We read the repository page but not its README text. |

An open-source tool has no vendor to sign a report. Shannon's README warns against running it on production, because it can create users and change application state.

## What do benchmarks say about automated tools?

Published benchmarks test narrow tasks. CyberGym (UC Berkeley) gives an agent a vulnerability description and a codebase and asks for a proof-of-concept, across 1,507 real vulnerabilities in 188 projects. Its abstract reports that top combinations achieved about a 20% success rate, for the 2025 models it tested. Cybench uses 40 professional-level capture-the-flag tasks.

Neither benchmark tests business-logic flaws, authorization flaws or your own application, so their scores show model skill but say nothing about a product's coverage of your app.

## How do you pick one?

Start from the question you need answered, then request a sample report and the retest term. The [sample report guide](/guides/judge-a-pentest-provider-by-its-sample-report/) shows how to check evidence quality. The [AI vs human pentesting](/guides/ai-vs-human-pentesting/) guide compares delivery types.

## Questions

### Can an automated tool replace a human pentest?

The vendors disagree and the evidence is mixed. Some products describe no human in the loop, and others add human review of each finding. Independent benchmarks test narrow tasks such as memory-safety bugs and capture-the-flag puzzles, and none of them test your app. Ask for a sample report and check what the tool says it could not test.

### Are open-source pentest agents safe to run on my systems?

Only on systems you own or have written permission to test, and not on production. Shannon's README says it can create users and change application state. Strix and PentestGPT carry similar authorized-use warnings. Review the license too: Shannon uses AGPL-3.0.

### Do automated tools find business logic flaws?

Some claim to. TurboPentest's page flags business-logic areas for manual follow-up and lists what automation could not test. Escape describes business-logic-aware DAST. Test the claim: ask for a finding where the tool proved a logic flaw with two accounts.

## Sources

- [XBOW pricing and product pages](https://xbow.com/pricing)
- [XBOW: The road to top 1](https://xbow.com/blog/top-1-how-xbow-did-it)
- [Horizon3.ai NodeZero pricing](https://horizon3.ai/use-case/pricing/)
- [Hadrian Nova](https://hadrian.io/products/nova)
- [Hadrian pricing](https://hadrian.io/pricing)
- [Cobalt pricing](https://www.cobalt.io/pricing)
- [Aikido AI pentest](https://www.aikido.dev/attack/aipentest)
- [Aikido pricing](https://www.aikido.dev/pricing)
- [Parameter pentesting](https://www.parameter.ai/pentesting)
- [TurboPentest pricing](https://turbopentest.com/pricing)
- [Terra Security](https://www.terra.security/)
- [Escape pricing](https://escape.tech/pricing)
- [Escape AI pentesting](https://escape.tech/product/ai-pentesting)
- [Synack pricing](https://www.synack.com/pricing/)
- [HackerOne pentest product page](https://www.hackerone.com/product/pentest)
- [Shannon (Keygraph) repository](https://github.com/KeygraphHQ/shannon)
- [Strix repository](https://github.com/usestrix/strix)
- [PentestGPT repository](https://github.com/GreyDGL/PentestGPT)
- [PentAGI repository](https://github.com/vxcontrol/pentagi)
- [CyberGym (arXiv 2506.02548)](https://arxiv.org/abs/2506.02548)
- [Cybench (arXiv 2408.08926)](https://arxiv.org/abs/2408.08926)