The AI pentester that never guesses.
AIpento is an autonomous AI penetration testing platform from UneeOne that attacks your own staging app the way an AI-equipped attacker would, reports only the weaknesses it can reproduce, and fixes them with patches verified three times from a clean checkout.
Map and attack
It attacks your app the way an AI attacker would.
Quorum's agents map the running app, propose a test plan and chain weaknesses together. A strict-code Prover sends every request, only inside the scope you authorised, and only against a host you proved you control with a DNS record. Production credentials are never requested.
- live attack probes
- 216
- static checks in the readiness scan
- 449
Prove or withdraw
AI proposes. Strict code decides.
Of the seven Quorum roles, three are AI and four are strict code, and not one marks its own work. A finding is kept only when the recorded request and response show the weakness. Anything that does not reproduce is withdrawn and never reported.
- Quorum roles
- 7
- unreproduced findings reported
- 0
Fix and verify
Fixed means your tests pass. Three times.
AutoFix writes a patch for every finding. It counts as fixed only when your own test suite passes from three clean checkouts and the original weakness no longer reproduces. You receive the pentest report, an executive brief and an attestation letter, mapped to compliance controls.
- clean-checkout test runs per fix
- 3
- compliance frameworks mapped
- 4
Operate
Command CenterAttack surfaceEngagementsFindings 2RemediationAssure
ReportsComplianceEngine · 216 probes- mappersurface map: 41 endpoints, 6 sign-in points, 2 LLM features
- plannerproposes: read another account's invoice via /api/invoices/{id}
- proverprobe AC-07: GET /api/invoices/1188 as member B -> 200, member A's data
- adjudicatorreproduced from a clean session -> CONFIRMED, high
- plannerproposes: SQL injection on /search?q=
- proverprobe IN-03: no differential response across 14 payloads
- adjudicatordid not reproduce -> WITHDRAWN, reason kept in the ledger
- autofixpatch written; your suite passes three times from clean checkouts
- scribeevidence: request + response, OWASP A01, ISO 27001 A.8.3, SOC 2 CC6.1
Quorum
- Mapper
- Planner
- Chain-Reasoner
- Orchestrator
- Prover
- Adjudicator
- Scribe
AI proposes Code decides
live attack probes
Counted from the probe catalogue in code, across 26 attack vectors: web applications, APIs, and AI and language-model features.static checks in the readiness scan
Counted from the code, across 21 lenses covering security, secrets, dependencies, infrastructure, reliability, release, AI agent tools and language-specific rules.kinds of leaked secret detected
API keys, tokens and passwords written into code or configuration. A secret's value is never reproduced in a finding: only its class, length and that it must be rotated.Quorum roles
Three AI models propose; four strict-code components orchestrate, execute, adjudicate and record.clean-checkout test runs per fix
AutoFix accepts a patch only when your own suite passes three times, each from a fresh checkout, and the original weakness no longer reproduces.unreproduced findings reported
A finding that cannot be reproduced against your running app is withdrawn, not reported with a lower score.compliance frameworks mapped
Findings and evidence are mapped to SOC 2, ISO 27001, PCI DSS and OWASP ASVS controls. Mapped evidence, not a certification.The problem
Why teams choose AIpento
AIpento sets autonomous agents on a running application you own and have authorised in writing: they map the attack surface, plan the tests and chain small weaknesses into real attack paths. Strict code, not an AI model, runs every request and decides what counts, so a finding that cannot be reproduced is withdrawn and every finding that remains carries the exact request and response that proved it. Alongside the live test, a static readiness scan reads the code without running it, and AutoFix writes patches that are accepted only when your own test suite passes three times.
Attackers now point AI at applications: it maps the surface and chains small weaknesses into a breach in minutes, continuously, at almost no cost.
A traditional penetration test takes weeks to book and describes the app as it was on the days it was tested; by the next release parts of it are already out of date.
Scanners work through a list of known tests and hand back long reports full of maybes, so teams spend days separating real findings from false ones and learn to distrust all of them.
An AI pentest that lets a model decide what is vulnerable produces confident mistakes; a system that guesses is worse than no answer at all.
Fixes are often accepted after one green test run, which can rest on leftover data or test order rather than a real repair.
How it works
How AIpento works, step by step
Connect
Grant read-only access to the repository for code scans, and name a staging address for the live test. Production credentials are never requested.
Prove ownership and authorise
A live test runs only against a host you have proved you control with a DNS TXT record, under a written authorisation that names it. Without that record the test does not start.
Scan the code
The readiness scan reads the repository line by line without executing any of it, and ranks every finding by severity with the file, the line and why it matters.
Map, plan and test
Quorum's agents map the running app, propose a test plan and suggest attack chains. A strict-code Prover sends every request, inside the authorised scope only.
Reproduce or withdraw
A finding is kept only when the recorded request and response show the weakness. Anything that does not reproduce is withdrawn and never reported.
Fix and verify
AutoFix writes a patch per finding. It counts as fixed only when your test suite passes three times from three clean checkouts and the original weakness no longer reproduces.
Report and retest
You receive the pentest report, an executive brief and an attestation letter, with findings mapped to compliance controls. Retests show what is new, what came back and what is still open.
Quorum
Why you can trust an AI pentest
Quorum is the autonomous multi-agent engine inside AIpento, built so that no single part is ever simply believed. Of its seven roles, three are AI models that propose and four are strict code that decides, and no role marks its own work. Agents propose. Strict code decides.
Coverage
What it checks, in the words your auditor uses
- Access controlReading or changing another account's data, acting above your role, weak sessions and tokens, including authenticated object-level and privilege-escalation tests when credentials are supplied.OWASP A01, A07 / API1, API3, API5
- InjectionSQL, NoSQL, command, template and XML injection, and cross-site scripting, including blind and time-based variants.OWASP A03
- Server-side request forgeryGetting your server to fetch an address an attacker chooses, including internal addresses, with blind cases confirmed through an out-of-band callback server.OWASP A10
- Misconfiguration and exposureMissing security headers, exposed files and backups, open admin panels and debug endpoints.OWASP A05 / API8
- API abuseMass assignment, missing rate limits, unrestricted resource consumption, sensitive business flows, improper inventory and unsafe use of third-party APIs.OWASP API Security Top 10
- AI features in your appDirect and indirect prompt injection, system-prompt leakage, sensitive data disclosure, improper output handling, excessive agency, poisoning, cross-tenant retrieval and unbounded consumption.OWASP Top 10 for LLM Applications
- AI agent tools and MCP serversThe code that gives AI agents their tools, checked for actions they should never be able to take.SAST / OWASP LLM06
- Leaked secrets204 kinds of API key, token and password written into code or configuration.Secrets detection
- Vulnerable packagesnpm and PyPI dependencies checked against the public OSV vulnerability database.SCA / OWASP A06
- Infrastructure as codeDockerfiles, Compose files, GitHub Actions workflows, Terraform and Kubernetes manifests.IaC scanning
- Source code securityLanguage-specific rules for JavaScript and TypeScript, Python, Go, PHP, Java, C#, Ruby, Rust, Kotlin, Swift and C/C++, read without being run.SAST
- Reliability and releaseMissing tests, swallowed errors, calls with no timeout, queries with no limit, missing migrations, missing lockfiles and loosely pinned versions.Code quality
Safety
How testing stays safe and in scope
- Live tests run only against a staging host you have proved you control and authorised in writing. No authorisation, no test.
- Production credentials are never requested, and nothing outside the scope you named is ever touched.
- No AI model ever sends a request to your app. Models propose; strict code is the only part that touches the target.
- Testing is bounded and non-destructive by default. Destructive probes run only with your explicit authorisation.
Deliverables
What you receive
- Pentest report with cover, executive summary, scope, methodology, every finding with CVSS v3.1 vector, the request and response that proved it, remediation, and an appendix of what was not tested and why
- Readiness report from the code scan: a verdict, findings by severity with file, line and why it matters, and a "needs your confirmation" section
- Executive brief: one page with the verdict, the three risks that matter and the recommendation
- Attestation letter: a signed statement of the test performed and its result, for your auditor
- Evidence Pack mapped to SOC 2, ISO 27001, PCI DSS and OWASP ASVS, with untested controls marked as not tested
- Verified patches from AutoFix, each with its three clean-checkout test results
- Retest results showing what is new, what came back and what is still open
- SARIF export of findings for your development tooling
Every feature
Everything AIpento does
6 areas, 28 capabilities, each with what it does for you.
01
AI Penetration Testing
An attacker's method, at the pace of your releases.
Autonomous agents test the running staging app you authorise, the way an AI-equipped attacker would, and keep only what they can prove.
- Proof of ownershipNothing is ever tested that you did not name, and your auditor can see the authorisation on file.
- Scope enforced in the request layerScope is a technical boundary, not a screen that the test ignores at runtime.
- 216-probe engineDepth where credibility is decided: access control, injection and AI features.
- Attack chainingYou see how small weaknesses combine into a real breach, not a flat list.
- Reproduce or withdrawNo false-positive noise to triage.
- Live attack consoleYou see the work being done, not a progress bar.
02
Readiness Scan
Your code, read line by line, never run.
A static readiness gate that reads the repository without executing it and gives the release one clear verdict.
- 449 static checksOne scan covers what usually takes several separate tools.
- Never executes your codeScanning is safe by construction.
- Confidence tiersQuestions that depend on intent go to a person instead of inflating the score.
- Every check can failA check that cannot fail cannot ship, so a clean result means something.
- Secrets never echoedThe report itself does not become a new leak.
- Scan from CISecurity checks run on every change, inside the tools your engineers already use.
03
AutoFix
Fixes you can trust, verified three times.
The fix squad: AutoFix writes a patch for a finding and accepts it only when the evidence says the weakness is gone and nothing else broke.
- Verified thriceSeparates a real fix from a lucky run that leaned on leftover data or test order.
- Fix certificateClosed means closed, with the reasons written down.
- Fix allClear a backlog without opening findings one by one.
- Model-agnosticThe verifier, not the model, decides whether a patch is accepted.
04
Reports and Evidence
A deliverable your auditor can check.
Every engagement produces reports built from the same findings the console showed, never a separate hand-written document that can drift.
- Pentest reportThe limits are stated, which is what makes the rest trustworthy.
- Executive briefGives the board the answer without the detail.
- Attestation letterSomething concrete to hand to an auditor or a customer's security review.
- Evidence PackAudit-ready evidence that never overstates coverage. Mapped evidence, not a certification.
05
Continuous Testing
Tested twelve times a year, not once.
An app you shipped to in March is not the app that was tested in January. Scheduled retests keep the answer current.
- Monthly retestThe report describes the app you run today.
- DriftRegressions surface before an attacker finds them.
- IntegrationsFindings land where your team already works.
06
Enterprise and Enclave
For many apps, strict access rules, or code that cannot leave your network.
The same engine runs as a hosted service or as AIpento Enclave, a self-hosted appliance inside your own network.
- AIpento EnclaveOn a default install, nothing leaves your premises.
- Outbound allow-list and transfer ledgerYour security reviewer can see exactly what crossed the boundary.
- Single sign-on and two-factorAccess is governed the same way as the rest of your estate.
- Tamper-evident audit logThe record lives somewhere an administrator here does not control.
- Tenant isolation by structureThere is no code path where one customer sees another's data.
Who it is for
Built for teams like these
- Teams shipping apps built with AI toolsCode from Lovable, Cursor, Replit, Bolt or Claude Code arrives fast; AIpento finds the failures that recur in it and fixes them before launch.
- Start-ups facing a customer security review or funding roundA one-time engagement delivers a proven report and an attestation letter to hand over, without weeks of booking.
- Engineering teams that release oftenMonthly retests and CI scans keep the security answer current with every release instead of once a year.
- Products with AI features or agentsDedicated probes target prompt injection, data leakage, excessive agency and the rest of the OWASP LLM Top 10, and the scan checks the code behind agent tools and MCP servers.
- Compliance and GRC leadsFindings arrive mapped to SOC 2, ISO 27001, PCI DSS and OWASP ASVS, with untested controls marked as not tested.
- Regulated and security-sensitive organisationsAIpento Enclave runs inside your own network with outbound traffic off by default.
FAQ
AIpento, answered
What is AIpento?
AIpento is an autonomous AI penetration testing platform made by UneeOne. It maps and attacks a staging copy of your app that you have authorised, reports only the weaknesses it can reproduce with the exact request and response, scans your code without running it, and fixes what it finds with patches verified three times from a clean checkout.
What is AI penetration testing?
AI penetration testing uses AI agents to do the work of a human penetration tester: map the attack surface, plan the tests, run them and chain small weaknesses into real attack paths. Done well, it brings the depth of a pentest to every release; done badly, it produces confident mistakes, which is why AIpento withdraws anything it cannot prove.
Is it safe to run AIpento against my app?
Yes, because it only tests a staging host you have proved you own and authorised in writing, and never production. Scope is enforced technically, testing is bounded and non-destructive by default, and no AI model ever sends a request to your app.
Does AIpento test production systems?
No. AIpento never asks for production credentials and only tests a staging address you own, named in a written authorisation and proved with a DNS record.
Does AIpento report false positives?
AIpento is designed not to: a live-test finding reaches you only if it was reproduced against your running app, with the request and response attached. Anything that does not reproduce is withdrawn rather than reported with a lower score, and code-scan items that depend on intent are marked "needs your confirmation" and never decide the verdict alone.
How is AIpento different from a vulnerability scanner?
A scanner works through a list of known tests and reports everything that might be wrong. AIpento's agents reason about your specific app, chain weaknesses into attack paths, and report only what strict code has reproduced.
How is AIpento different from SAST?
SAST reads code but cannot see how the running app behaves. AIpento includes a static readiness scan of 449 checks and also attacks the running app, so it finds problems that only exist when the pieces run together.
How is AIpento different from a manual penetration test?
A manual pentest takes weeks to book and describes the app as it was on the days it was tested. AIpento applies the same method of mapping, planning, testing and chaining at the pace of your releases, with monthly retests, and every finding carries its proof.
What is Quorum?
Quorum is AIpento's multi-agent engine of seven roles: three AI models that map, plan and chain, and four strict-code components that orchestrate, execute, adjudicate and record. Agents propose; strict code decides, and no role marks its own work.
What do I get at the end of an engagement?
You get a pentest report with the evidence for every finding, an executive brief, an attestation letter for your auditor, and an Evidence Pack mapped to SOC 2, ISO 27001, PCI DSS and OWASP ASVS. The report also lists what was not tested and why.
Does AIpento fix the issues it finds?
Yes. AutoFix writes a patch for each finding and accepts it only when your test suite passes three times from three clean checkouts, the original weakness no longer reproduces, and no new High or Critical weakness appears; Fix all queues a fix for every open finding in one step.
Can AIpento run inside our own network?
Yes. AIpento Enclave is a self-hosted appliance that runs inside your network with outbound traffic off by default and a licence checked offline, so on a default install nothing leaves your premises.
Does AIpento make us SOC 2 or ISO 27001 compliant?
No. AIpento supplies mapped evidence, not certification; compliance needs an independent auditor. A control no check covers is recorded as not tested, never as a pass.
Who makes AIpento?
AIpento is made by UneeOne, an AI security company. UneeOne also makes UneeGuard AI for AI data and agent security.
See AIpento on your own app.
Connect a repository for a readiness scan in minutes, or prove a staging host and run a live pentest.