In short
- XBOW describes itself as an autonomous offensive security platform that attacks applications and APIs and reports only validated findings, each with the working exploit, the request and the response.
- In a post dated 24 June 2025, XBOW said it had reached the top spot on HackerOne's US leaderboard after submitting nearly 1,060 vulnerabilities. Its security team reviewed the findings before submission.
- XBOW publishes no price list. Its pricing page says pricing is usage-based and scoped to your environment, with a quote or a purchase through a cloud marketplace.
- XBOW is built for security teams with many applications. A solo founder is not its buyer, and that is no criticism of the product.
- What XBOW proved matters to everyone: attacking a web app no longer needs a person, so every public app gets tested by someone. Test yours first.
On this page
What is XBOW?
XBOW is a company that builds an autonomous AI pentester. Its site describes an "autonomous offensive security platform" that attacks applications and APIs "the way an elite attacker does" and reports only what it actually breaks into. There is no human tester in the loop for the attack itself. AI agents explore the target, form ideas about what might be broken, try them, and keep what worked.
The part that sets XBOW's pitch apart from older scanners is validation. The company says findings in its generally available classes are confirmed by a deterministic check, and that each finding ships with the working exploit, the request and response, and steps to reproduce it. After a fix, XBOW tests again to confirm the exploit no longer works, and its site says those retests are free and unlimited.
XBOW is well funded. On 18 March 2026 it announced a $120 million Series C led by DFJ Growth and Northzone, at a valuation it put at over $1 billion. Its home page says more than 200 security teams use the platform.
The HackerOne ranking, from the primary source
HackerOne is a bug bounty platform: companies invite researchers to find flaws and pay for valid reports. Researchers are ranked by reputation on leaderboards. In a blog post dated 24 June 2025, XBOW wrote: "For the first time in bug bounty history, an autonomous penetration tester has reached the top spot on the US leaderboard." The same post says XBOW submitted nearly 1,060 vulnerabilities.
Two details from that post are worth keeping, because press summaries often drop them:
- The leaderboard named in the post is the US one. Some articles describe it as a global ranking. XBOW's home page today says it ranked first on HackerOne, above every human researcher, without naming the leaderboard. When sources differ, the dated post is the precise one.
- Humans reviewed before submission. XBOW says its security team reviewed all findings before they were sent, to comply with HackerOne's policy on automated tools. The agents found and validated. People pressed send.
The post also explains how the system avoids false alarms: "validators", automated reviewers that confirm each finding. For cross-site scripting, for example, a headless browser visits the target to confirm the script actually ran. That design choice, proof before report, is the reason the result was possible at all. A bug bounty program rejects reports that cannot be reproduced.
Who XBOW sells to, and how it is priced
XBOW's site speaks to security teams. Its testimonials come from security leaders, and it presents a case of a very large company that assessed 180 applications in a month. The pricing page lists no plan and no number. It says pricing is "scoped to your environment" and usage-based, covers "a single application to thousands", and can be bought on quote or through the AWS, Google, Oracle and Microsoft marketplaces.
| Question | What XBOW's site says |
|---|---|
| What does it test? | Applications and APIs you expose, continuously or on demand |
| Who runs it? | Security teams. The site says more than 200 use it |
| How is a finding proven? | A deterministic check for generally available classes. The working exploit, request and response are attached |
| What about retests? | Free and unlimited, after a fix is deployed |
| What does it cost? | No public price. Usage-based, on quote or through a cloud marketplace |
That is a normal way to sell to large companies. It also means a founder with one Lovable app and no security team is not the intended buyer: there is no self-serve plan to pay for with a card on that page.
What an autonomous AI pentester means for a small team
Three things change for you, even if you never buy one.
- Attack got cheap. Finding a table left open or a route with no session check used to take a person's afternoon. Agents do it at scale. The assumption that a small app is too obscure to be tested no longer holds.
- Proof became the standard. XBOW, and the tools that followed, report the exploit that worked. A list of maybes from a scanner now looks dated. Ask any tool you consider for the request and the response behind each finding.
- The pentest stopped being yearly. When a test runs in hours, it can run after each release. The guide on automated pentests covers what that rhythm looks like.
The apps most exposed to cheap attacks are the ones shipped fastest. The public number for AI-built apps is CVE-2025-48757: missing or weak row level security in 170 of 1,645 Lovable apps tested, about one in ten. Finding them took a script.
Three checks to run on your own app today
You do not need an AI agent for the first pass. These are the checks an agent, or a bored stranger, runs first. Use your own app and test accounts you created.
- 1
Ask your database without signing in
On Supabase, request one of your tables with only the public key, as in the command below. A private table should return an empty list or a permission error. On Firebase, run the Firebase rules checker on your app's address.
- 2
Ask for another account's record
Create two accounts. As the first, create a record and note its id. As the second, open the same address or call the same API route with that id. You should get a 403 or a 404. Getting the record is the flaw called IDOR.
- 3
Search what your site sends to every visitor
Open your site, open the developer tools, and search the loaded JavaScript for
service_role,sk_liveandsk-. A match means a secret key is public. Rotate it first, then move the call to your server. See what to do about a leaked API key.
# 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"alter table public.your_table enable row level security;
create policy "Owners read their rows"
on public.your_table for select
to authenticated
using ( (select auth.uid()) = user_id );XBOW and Keybrum: two different designs
We build Keybrum, so read this section as the view of a much smaller product. The two share one belief: a finding without proof is noise. They differ in almost everything else.
| XBOW | Keybrum | |
|---|---|---|
| How it attacks | Autonomous AI agents that explore the target and adapt | A fixed library of attacks, each written and proven in advance, run by code |
| What it is built for | Any application or API a large organization exposes | Apps built with Lovable, Bolt, v0, Replit, Cursor, Claude Code and Base44, on Supabase, Firebase or their own backend |
| Buyer | Security teams with a portfolio of applications | The founder or the agency that ships the app |
| Price | Usage-based, on quote | €49 for a one-time report. Plans from €29 a month excluding VAT |
| After the fix | Retests confirm the exploit no longer works | The same attack runs again, on demand, by your coding agent or on the next deploy. A finding closes when the attack fails |
The trade is clear. An agent that explores can find a flaw nobody wrote an attack for, including logic specific to your product. A fixed library cannot. In return, a library gives the same run every time: the same attacks, the same cost, a result you can compare from one deploy to the next, and no model deciding what to try against your production data.
For the stacks AI builders generate, a library goes a long way, because these apps fail in the same few places. If your app is large, custom and carries regulated data, an autonomous platform or a manual pentest sees more. The buyer's list of AI pentest tools places both next to the other products in this field.
What Keybrum tests, and what it does not
On a Supabase app, Keybrum reads tables with the public key, tries to insert rows and to change 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 an app with its own backend it calls API routes as a stranger and as two accounts it creates. It looks for secret keys in the bundle and confirms each with a read-only call, checks for exposed .env files and source maps, AI endpoints anyone can spend, unsigned Stripe webhooks, and known Next.js and React flaws proven by behaviour.
It does not explore your app the way an autonomous agent does, so it will not find a flaw outside its library. It does not test mobile apps or internal networks. It never tests delete access, never runs remote code execution or denial-of-service payloads, and only attacks an app whose owner proved ownership. A clean run reads as a list of attacks that failed on a date, with what was not tested next to it.
Each finding carries the request and the response, with personal data redacted, and a fix your coding agent reads through Keybrum's MCP server. A one-time report on one app is €49, with retests of its findings for 7 days. Plans attack every deploy: Maker at €29 a month for 5 apps, Team at €99 a month for 25 apps with your brand on reports, both excluding VAT.
Questions
What is XBOW AI?
XBOW is a company whose platform uses autonomous AI agents to attack web applications and APIs, the way a penetration tester would. It reports only findings it could validate, each with the working exploit and the request and response. It is sold to security teams.
Did XBOW really rank first on HackerOne?
XBOW's own post of 24 June 2025 says it reached the top spot on HackerOne's US leaderboard, after submitting nearly 1,060 vulnerabilities. Some press articles describe the ranking as global. The post also says XBOW's security team reviewed the findings before submission.
How much does XBOW cost?
XBOW does not publish prices. Its pricing page says pricing is usage-based and scoped to your environment, and that you can request a quote or buy through the AWS, Google, Oracle or Microsoft marketplaces.
Is XBOW a good fit for a startup or a solo founder?
It is built and sold for security teams that cover many applications. A solo founder will find no self-serve plan on its pricing page. For one app, start with the three checks on this page, then a proof-based automated test priced for one app. See pentests for startups.
Will AI pentesters replace human pentesters?
They already do part of the work: finding and proving common flaws at a volume no person matches. Human testers still lead on business rules, on chains of small flaws, and on reports an auditor signs. Most teams will use both.
What is the difference between XBOW and Keybrum?
XBOW uses autonomous agents that explore any application, for security teams, priced on quote. Keybrum runs a fixed library of proven attacks for apps built with AI tools on Supabase, Firebase or their own backend, for founders, from €49. XBOW can find what nobody anticipated. Keybrum gives the same run on every deploy.
Sources
- 1.XBOW home page, XBOW
- 2.Blog post of 24 June 2025 on reaching the top of HackerOne's US leaderboard, XBOW
- 3.XBOW raises $120M to scale its autonomous hacker, XBOW
- 4.Pricing, XBOW
- 5.Row Level Security, Supabase Docs
- 6.CVE-2025-48757, Matt Palmer