In short
- Published prices range widely. Burp AT starts at $19 per user a month on top of a Burp Suite Professional licence. Aikido lists a typical AI pentest at €3,500. NodeZero lists web app pentesting at $10,000 per app, per test. XBOW and RunSybil publish no price.
- Every serious tool now claims proof per finding. Ask to see one finding from a sample report: the request, the response, and the steps to replay it.
- Open source agents such as Shannon and Strix cost only model usage, and they need you to run them. Shannon's own README says not to run it against production systems.
- An agent that explores can find flaws nobody anticipated. A fixed library finds only what it has attacks for, and gives the same run on every deploy. Know which one you are buying.
- Before any tool touches your app, check who it lets attack what. A tool that attacks any address typed into it can be pointed at you by anyone.
On this page
The four kinds of AI pentest tools in 2026
The label AI pentest tool now covers products that have little in common. This page is a buyer's list. For what AI agents can and cannot do in a pentest, and why proof matters, read the guide on AI penetration testing first. Here we sort the products, quote what each vendor publishes about itself, and give a method to choose.
| Kind | How it works | Who buys it | Examples |
|---|---|---|---|
| Autonomous platform | AI agents attack your applications continuously or on demand, and report validated findings | Security teams with many applications | XBOW, RunSybil, NodeZero |
| AI pentest per assessment | You order a test of one app. Agents run it in hours and you get a report an auditor can read | Teams that need a pentest report for a customer or SOC 2 | Aikido |
| Assistant for human testers | AI inside the tester's own toolkit. The person stays in charge | Pentesters and bug bounty hunters | Burp AT |
| Open source agent | You run the agent yourself, with your own model API key | Developers and security engineers who can operate it | Shannon, Strix |
A fifth kind sits next to these and is often listed with them: the proof-based automated pentest that runs a fixed library of attacks written in advance. It uses no model to decide what to attack. Keybrum, which we build, is one of these. We explain below why that matters and when it is the weaker choice.
Eight tools, and what each says about itself
Everything in this section comes from the vendor's own site or repository, read on 13 November 2026. We have not benchmarked these products against each other, and we give no ranking. A price appears only when the vendor publishes one.
XBOW
XBOW calls itself an autonomous offensive security platform that attacks applications and APIs and reports only what it breaks into. Its site says findings in generally available classes are confirmed by a deterministic check and come with the working exploit, and that retests are free and unlimited. It sells to security teams. No public price: the pricing page says usage-based, on quote or through cloud marketplaces. Our page on XBOW covers its HackerOne result.
Aikido AI Pentesting
Aikido is a security platform for code and cloud that also sells AI pentests. Its site says findings are reported only after successful exploitation against the live target, that source code access is optional and "significantly improves results", and that a report for SOC 2 or ISO 27001 arrives within hours. Published prices: a typical pentest at €3,500 per assessment for one application and its primary APIs, with free retesting for up to 6 months, and a rightsized pentest from €45 to more than €25,000 depending on the app. Both carry the line "No High or Critical Finding = Don't Pay". See Keybrum vs Aikido.
Burp AT (PortSwigger)
PortSwigger, the maker of Burp Suite, describes Burp AT as "agentic AI that extends human-led pentesting". It is in public beta for Burp Suite Professional users. The tester picks a mode (Manual, Smart or Autonomous) and sets approval rules, and PortSwigger says scope is enforced by Burp's tooling layer instead of being left to the model. Published prices for individuals: $19, $89 or $249 per user a month depending on usage, each on top of a Burp Suite Professional licence. This is a tool for someone who already knows how to test. See Burp Suite alternatives.
NodeZero (Horizon3.ai)
NodeZero comes from infrastructure testing: internal networks, external assets, cloud and Kubernetes. Horizon3.ai says it provides proof of the exploit, its impact and mitigation advice. It is one of the few vendors here with a public price list: packages from $50 to $85 per asset a year with unlimited testing, or $15 per asset per test. Web application pentesting is a separate offer, listed at $10,000 per app, per test, or custom pricing for continuous testing.
RunSybil
RunSybil describes an AI-powered offensive security platform that continuously tests applications and infrastructure, maps code, APIs and cloud, and chains flaws into attack paths. It says it proves whether what your other tools found is actually exploitable. Its site addresses security teams with critical systems. No public price.
Shannon (Keygraph)
Shannon is an open source AI pentester for web applications and APIs, under the AGPL-3.0 licence. The open source edition is white-box: it reads your source code, then attacks the running app. Its rule is "No exploit, no report". The README says a full run takes roughly 1 to 1.5 hours and that you pay your own model costs. It also says plainly: "Do not run Shannon against production systems." A paid Pro edition is listed at $50 per active developer a month.
Strix
Strix calls itself "the open-source AI pentesting tool", under the Apache 2.0 licence. Its agents run your code dynamically and validate findings with proofs of concept. To run it you need Docker and an API key from a model provider. The README marks it for authorized use only, on systems you own or have explicit written permission to test. A hosted version, Strix Cloud, exists, and the README gives no price for it.
Keybrum
Keybrum attacks a live app with a fixed library of attacks built for apps made with Lovable, Bolt, v0, Replit, Cursor, Claude Code and Base44, on Supabase, Firebase or their own backend. No model chooses what to send to your app. Every finding is the request that worked. Prices: €49 for a one-time report on one app, plans from €29 a month excluding VAT. Its limits are in the last section of this page.
Published prices, side by side
| Tool | Published price | What you must bring |
|---|---|---|
| XBOW | None. Usage-based, on quote | A security team to buy and run it |
| Aikido AI Pentesting | €3,500 per typical pentest. Rightsized from €45 | Scope and test credentials. A repository improves results |
| Burp AT | $19, $89 or $249 per user a month | A Burp Suite Professional licence and testing skills |
| NodeZero | $10,000 per web app, per test. Infrastructure from $15 per asset per test | A scope of assets |
| RunSybil | None | A demo request |
| Shannon | Open source: your model costs. Pro: $50 per active developer a month | Source code, a test environment, an operator |
| Strix | Open source: your model costs | Docker, a model API key, an operator |
| Keybrum | €49 once, or from €29 a month excluding VAT | The app's address and proof of ownership |
These prices buy different things and should not be read as a ranking. A €3,500 pentest with an attestation letter and a $19 assistant for a tester are not substitutes. Use the table to rule out what does not fit your size.
How to choose: four questions for any vendor
- 1
Is there proof for every finding?
Ask for a sample report and open one finding. You should see the exact request, the response, the account it ran as, and steps to replay it. A severity score and a paragraph of text are an opinion. Our pentest report sample shows what proof looks like.
- 2
Is it safe on a live app?
Ask what the tool will never do. Does it write to your database? Delete anything? Send destructive or resource-heavy payloads? Is there a stop button? An agent that explores freely on production data needs strict limits. Shannon's maintainers tell you to use staging with disposable data, which is the honest answer for that design.
- 3
Does it check ownership?
Ask what stops someone else from pointing the tool at your app. A hosted product should make the customer prove control of the target, by a DNS record, a file or a tag, before anything is sent. Open source agents leave that responsibility with the person who runs them.
- 4
What happens after the fix?
A finding is closed when the same attack fails. A ticket marked done proves nothing. Ask whether retests are included, for how long, and whether the test can run on every deploy. A report that is right for one day has little value on a product that ships daily.
Which kind fits you
| Your situation | Start with |
|---|---|
| Solo founder, one app built with an AI tool, no repository habits | The self-check below, then a fixed-library test priced per app |
| Developer comfortable with Docker, app in a repository, a staging copy | An open source agent on staging |
| A customer asks for a pentest report this month | An AI pentest per assessment with a formal report, or a manual test. See pentest cost |
| You are a pentester | An assistant inside your own toolkit |
| Security team with dozens of applications | An autonomous platform, on a trial scope |
Many teams end up with two: something that runs on every deploy, and a deeper test once or twice a year. The guide on automated pentests explains how the two fit together.
Before you buy: check the three usual leaks yourself
Whatever tool you choose, it will look at these first. So can you, on your own app, in ten minutes.
- Your database, without signing in. On Supabase, request a table with only the public key (command below). A private table should return an empty list or an error. The fix is row level security.
- Another account's record. Create two accounts. As the second, ask for a record the first created. Anything but a 403 or a 404 is a finding. See IDOR.
- Secret keys in the bundle. Search your site's loaded JavaScript for
service_role,sk_liveandsk-. See API keys exposed in the frontend. - Your AI route. Call the endpoint that forwards prompts to a model without a session. If it answers, anyone spends your credits. See unprotected AI endpoints.
# Your own project only. No sign-in, just the public key.
curl "https://YOUR_PROJECT_REF.supabase.co/rest/v1/YOUR_TABLE?select=*&limit=5" \
-H "apikey: YOUR_PUBLIC_KEY"Where Keybrum sits, honestly
Keybrum belongs on this list with a caveat: the attacks are not chosen by an AI. Each one was written, tested and bounded by us in advance. On Supabase it reads tables with the public key, tries inserts and updates on other people's rows, calls database functions, lists storage buckets and subscribes to realtime channels, as a stranger and as two accounts. On Firebase it tests Firestore, Realtime Database and Storage rules. On your own backend it calls API routes as a stranger and as two accounts it creates. It confirms leaked keys with read-only calls, and checks for exposed .env files, AI endpoints anyone can spend, unsigned Stripe webhooks, and known Next.js and React flaws proven by behaviour.
Where AI comes in is the fix. Your coding agent (Cursor, Claude Code) reads each finding and its fix through Keybrum's MCP server, changes the code, and triggers the retest. The finding closes when the attack fails.
What it does not do: explore your app for flaws outside its library, test business rules specific to your product, test mobile apps or internal networks, or produce a report signed by a human tester. It never tests delete access and never runs remote code execution or denial-of-service payloads. It only attacks an app whose owner proved ownership. If you need unknown flaws found in a large custom app, the autonomous platforms above or a manual pentest see more.
Questions
What is the best AI pentest tool?
There is no single best one, because the products are built for different buyers. Security teams with many applications look at autonomous platforms such as XBOW or RunSybil. Pentesters look at Burp AT. Developers with a staging environment can run Shannon or Strix. A founder with one AI-built app is better served by a test priced per app.
Are there free AI pentesting tools?
Shannon and Strix are open source. The software costs nothing, and you pay the model provider for the tokens a run uses. You also need the skills to run them and a safe environment. Shannon's README advises against running it on production systems.
How much does AI pentesting cost?
Published prices on 13 November 2026 ranged from $19 per user a month for Burp AT's lightest tier, on top of a Burp Suite Professional licence, to €3,500 for Aikido's typical pentest and $10,000 per app, per test, for NodeZero's web app offer. XBOW and RunSybil quote on request. Keybrum's one-time report is €49.
Can an AI pentest replace a manual penetration test?
For common, well-understood flaws, increasingly yes, and it can run far more often. For business rules, chained attacks and a report signed by an accredited tester, a person is still needed. Some auditors accept AI pentest reports and some ask for a human-led test, so ask yours first.
Is AI pentesting safe to run on production?
It depends on the tool's limits. Ask what it will never do: write, delete, send heavy payloads. Hosted products usually run on production with guardrails and a way to stop. Some open source agents tell you to use staging. Never run any of them on an app you do not own.
Is Keybrum an AI pentest tool?
Partly. Keybrum's attacks are a fixed library run by code, with no model choosing what to send to your app. AI is used on the fixing side: your coding agent reads findings and fixes through Keybrum's MCP server and triggers retests.
Sources
- 1.XBOW home page, XBOW
- 2.Pricing, XBOW
- 3.AI Pentesting, Aikido Security
- 4.Pricing, Aikido Security
- 5.Burp AT, PortSwigger
- 6.Burp AT pricing, PortSwigger
- 7.NodeZero, Horizon3.ai
- 8.Pricing and packages, Horizon3.ai
- 9.RunSybil home page, RunSybil
- 10.Shannon README, Keygraph
- 11.Strix README, Strix