A database anyone can read
The browser talks straight to a hosted database. With row security off or rules left open, anyone with the public key can read or change every row.
We check row security, storage rules, and which keys ship to the browser.
Custom-built AI hackbots. Hands-on human testing.
Xarlo Security: penetration testing for web apps, APIs, mobile apps and AI-built software. AI wrote your code. We find what it missed.
Xarlo runs its own custom-built hacking agents alongside senior researchers who test by hand. The hackbots cover every route, role and release at machine speed. Our researchers direct them, chase the ideas no model would try, prove every finding, and help you fix it.
Fixed price. Nothing unverified in the report. Every fix retested.
Routes mapped214
Hackbot checks9,412
Researcher directives3
ReconHackbotsHuman testingReport
Critical 2
High 1
Medium 1
Unverified in report 0
- Vulnerabilities reported by our researchers
- 880+
- Engagements with hands-on human testing
- 100%
- Unverified findings in our reports
- 0
- Retest of every fix
- Free
Built with AI or built by hand, we test it
Fast to build. Fast to break.
AI app builders and coding assistants get a product live in days. They also repeat the same six security mistakes, and attackers know exactly where to look.
Secrets in the front end
Service keys, API tokens and webhook secrets end up in the JavaScript bundle that every visitor downloads.
We check every bundle, source map and env file, and prove each key live or dead.
No ownership checks
Generated endpoints fetch records by ID and never ask whether the record belongs to the person asking.
We check every route, role and object type, across real accounts.
Security that lives only in the UI
Admin buttons are hidden from normal users, but the API behind them answers anyone who calls it.
We check that the server enforces every action the UI hides.
AI features that take orders from strangers
Chatbots and agents follow instructions hidden in tickets, documents or web pages, then leak data through their tools.
We check prompt injection, tool permissions, and what the model can reach.
Packages that should not exist
Assistants suggest dependencies that are misspelled, abandoned, or registered by attackers after a model invented the name.
We check your full dependency tree, lockfiles and install scripts.
How exposed is your app?
Tick what applies. You get a rough exposure level and the engagement we would start with. It is a guide built on the patterns above, not a scan of your app.
40Elevated exposure
Where we would start
- AI app launch review
- Access-control testing across accounts
Your answers stay in this page. Nothing is sent anywhere.
Hackbots for breadth. Hackers for depth.
Xarlo hackbots are hacking agents we build and run on our own harness. They produce thousands of signals. Our researchers steer them with new instructions, test by hand alongside them, throw out the noise and prove the rest.
Recon hackbothosts, routes, versions
Access hackbotevery role on every route
Fuzzing hackbotinputs, parsers, uploads
AI-feature hackbotprompts, tools, memory
Manual testingresearchers, by hand
Researcher review
Critical
High
Medium
- Signals from hackbots
- 12,480
- Instructions from researchers
- 312
- Leads tested by hand
- 1,146
- Verified and reported
- 47
What our hackbots do
- Map every host, route, parameter and API version
- Run the full permission matrix: every role against every route
- Fuzz inputs, uploads and parsers at a pace no person can match
- Hit AI features with thousands of injection variants
- Take new instructions from a researcher in the middle of a run
- Re-run everything on each release so old issues stay fixed
What our researchers do
- Test by hand with open-source tools plus custom tools and scripts our researchers build
- Think like an attacker: which feature is worth breaking?
- Point the hackbots at short, critical ideas no model would try
- Find business-logic flaws and chain small issues into a real attack path
- Prove impact with a request you can replay, then write the fix
Inside the Xarlo harness
Our hackbots run on a harness we built ourselves. This is a replay of the first hour of an engagement: recon, the access matrix and the secret scanner side by side, with a researcher steering what runs next.
0:recon214 routes
1:access-matrix80 / 80 checks, 3 violations
allowed, correctdenied, correctviolation16 routes per role
2:secrets2 leads
[xarlo]0:recon1:access-matrix*2:secrets00:00:00
Every issue comes with proof
Six replays of how we work. A researcher tests by hand with curl, a script or a hackbot shows how far the flaw reaches, and every result gets a CVSS score.
- 1Baseline, by handA normal request, sent with curl.
- 2Attack, by handThe same request with one value changed.
- 3Scale, by script or hackbotHow far it reaches, then a score and a fix.
CVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:U/C:H/I:H/A:N
Mobile apps, taken apart
We pull the build from the store, decompile it, instrument it on rooted and jailbroken devices, and test the API underneath. Five replays of what that looks like on Android and iOS.
lab@android: ~/acme-walletRooted Android device
Medium 6.5
Small flaws add up to a breach
Scanners and AI-only tools report each issue on its own and rate it low. A researcher asks what they add up to. Scroll to follow one chain from a public file to admin access.
-
Info on its own
Source map lists a legacy admin route
The public bundle ships a source map that names
/api/v1/admin/invite. -
Low on its own
The old v1 API is still deployed
v2 replaced it months ago, but v1 still answers on the same host.
-
Medium on its own
v1 invite skips the role check
Any member can send invites, not just admins.
-
The link that joins them
Invite trusts orgId from the request
So a free-trial account can invite itself as admin into any customer’s workspace.
Combined: admin access to any customer workspace
Four issues a scanner would file as noise. Together, one critical path we proved end to end.
High 8.8
CVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:U/C:H/I:H/A:H
app/api/invoices/[id]/route.ts
Fix written
Reviewed by a researcher
Pull request opened
Retest passed
We don’t just report it. We help fix it.
Most AI-built apps don’t have a security engineer on the team. So every finding comes with a fix written for your stack, and our fix sprints deliver the patch itself.
- Patches in your codebaseDelivered as pull requests, written in your framework’s idioms.
- Secure defaults so it stays fixedPolicies, middleware and tests that fail if the issue comes back.
- Prompts for your AI assistantFix instructions your team can hand to its AI coding assistant, reviewed by us after.
What we test
Nine kinds of engagement, all led by senior researchers and backed by our hackbots. Open one to see what we look at and what you get at the end.
What we look at
- Row security, database rules and storage buckets
- Keys, tokens and secrets in the front end
- Ownership checks on every generated route
- Sign-up, login, reset and session handling
What you get
A prioritized fix list your team or your AI assistant can apply, plus a retest.
Typical length
About 1 week
What we look at
- Direct and indirect prompt injection
- What each tool can reach, and on whose authority
- Leaks between users through context, memory or retrieval
- System prompt and secret exposure
What you get
Attack transcripts for every issue and concrete guardrails for your stack.
Typical length
1 to 2 weeks
What we look at
- Patches written in your framework and conventions
- Policies and middleware so the class of bug cannot return
- Tests that fail if an issue comes back
- A walkthrough with your engineers
What you get
Merged fixes and a clean retest letter.
Typical length
Days, not weeks
What we look at
- Every role against every endpoint, sideways and upward
- Login, reset, invites, MFA and sessions
- Whether one customer can read or change another’s data
- Payments, quotas and workflow abuse
What you get
A report with a CVSS 3.1 vector per issue, requests your engineers can replay, and fixes.
Typical length
2 to 3 weeks for a SaaS product
What we look at
- Object-level and function-level authorization on every route
- Old, undocumented and deprecated API versions
- Server-side requests toward internal services and cloud metadata
- Leaked keys, proven live or discarded, and over-broad roles
What you get
Issues mapped to the exact route and role, plus an inventory of every API surface we found.
Typical length
1 to 3 weeks
What we look at
- The shipped build, for secrets and hidden endpoints
- Runtime instrumentation past pinning and client-side checks
- Local storage, deep links and exported components
- The mobile API, which often lags behind the web one
What you get
Separate results per platform, with device-side and server-side causes kept apart.
Typical length
1 to 2 weeks per platform
What we look at
- Exposed services, remote-access gateways and forgotten hosts
- Vulnerable appliances, confirmed by testing rather than by banner
- Paths from a standard domain user to domain admin
- Whether segmentation holds between environments
What you get
An attack-path narrative with the one change that breaks each path.
Typical length
1 to 2 weeks
What we look at
- Goals set with you first, such as reaching production data
- Initial access through your real external attack surface
- What your monitoring caught, and how quickly
- A joint debrief with the people who defend you
What you get
A timeline of every action next to what your monitoring recorded.
Typical length
4 to 6 weeks
What we look at
- New features and routes as they ship
- Regression checks on every past issue
- New hosts and changes on your external surface
- The same assigned researcher every month
What you get
Issues as they are confirmed rather than at the end, plus a monthly summary for leadership.
Typical length
Monthly, cancel any time
Scanners, AI-only tools, and Xarlo
Automation finds the obvious. AI-only testing finds more, with more noise. Breaches in modern products come from logic, and logic needs a person who understands your product.
| Area | Automated scanner | AI-only testing | Xarlo: hackbots and hackers |
|---|---|---|---|
| Maps every route and role | Partly | Yes | Yes, with our hackbots |
| Hands-on manual testing | None | None | On every engagement, by senior researchers |
| Permission flaws between users | Not modeled | Flags guesses | Every lead proven or discarded by a researcher |
| Business logic | Out of reach | Shallow | Tested against how your product should work |
| Chained attacks | One request at a time | Rarely | Small flaws combined into full attack paths |
| Evidence | A signature match | Model output | A request you can replay yourself |
| Fixes | Generic advice | Generic advice | Patches and pull requests for your stack |
| Unverified items in the report | Many | Some | None |
How an engagement runs
Five steps from the first call to the retest. You work directly with the researchers testing your systems at every one of them.
Scope
1 to 3 daysWe agree on assets, test accounts and rules of engagement, then send a fixed price.
Recon
First days of testingHackbots map every host, route and role. A researcher builds the threat model.
Exploit
Most of the windowHackbots run the matrix at scale while researchers test by hand and take each lead to the end.
Report and fix
Within 5 working daysVerified issues only, each with a CVSS 3.1 vector, replayable proof and a patch.
Retest
Free, within 90 daysWe confirm each fix and issue an updated report and attestation letter.
Our practice
Built by senior offensive security researchers with more than 880 reported vulnerabilities between them. Three practice areas, each led by a senior researcher, and no work passed to junior testers.
Application and AI security
Web apps, APIs, AI-built products and LLM features: access control, identity, prompt injection and tenant isolation.
Mobile and attack surface
iOS and Android apps, runtime instrumentation, and finding the exposed assets you did not know you had.
Infrastructure and cloud
External and internal networks, domain environments, and identity and access in your cloud accounts.
Methods aligned with
- OWASP ASVS
- OWASP MASVS
- OWASP API Top 10
- OWASP Top 10 for LLM Apps
- PTES
- NIST SP 800-115
- CVSS 3.1
Researcher profiles, track records and client references are shared under NDA during scoping.
Questions teams ask us
Straight answers on how we test, what we touch, and how pricing works.
Something else on your mind?
Book a scoping call
Ask it on a 30-minute scoping call with a researcher.
Hacking agents we build and run ourselves, on our own harness. They map your app, run the permission matrix, fuzz inputs and attack AI features at machine speed, and they take instructions from the researcher leading your engagement.
No. Every engagement includes hands-on testing by senior researchers, using open-source tools plus custom tools and scripts we build ourselves. The hackbots cover breadth. People find the business-logic flaws and attack chains no model thinks to try.
Yes, and these apps are a large part of our work. We test the deployed app, its database rules and its API, and we can review the generated code if you share it.
It stays inside the engagement. We only test what is in scope, share results with your team alone, never use your data to train models, and delete it when the engagement ends unless you ask us to keep it.
We agree on rules of engagement first. Risky tests run on staging or on test accounts you create, and anything we change in production is reverted the same day.
Both. Every report includes specific fixes, and our fix sprints deliver patches as pull requests with a retest.
A fixed price per engagement, based on scope and agreed after one scoping call. No hourly billing.
Yes, before any access is granted. Researcher profiles and references are shared under NDA.
Tell us what you run. We’ll send a fixed quote.
No obligation and no sales sequence. A senior researcher reads every request.
- 1You send a short briefWhat you built, how, and what worries you.
- 2We have a 30-minute scoping callWith a researcher, not a salesperson.
- 3You get a fixed price and datesAccept it, change it, or walk away.
Prefer email? [email protected]
Your brief is ready
Copy it and send it to [email protected]. We reply within one working day.