NOT A SCANNER. PROOF, NOT PROBABILITIES.
If it does not reproduce, it is not a finding.
A scanner hands your reviewer a pile of maybe. NAXXAN is an autonomous red team that attacks your deployed surface, fires each exploit twice with a clean negative control, and hands your reviewer evidence — your reviewer can re-run it.
$995 · first five engagements only · report in 48h from the agreed start · full refund before execution
The reviewer wants evidence. Your tools hand them opinions.
One of these is probably true right now: an enterprise security review is asking for proof of your external posture, a release with real exposure ships next month, or an AI feature is about to go to production. And after a season of public reporting on autonomous agents, companies running them in production are fielding a harder question from above: how do we know what ours do when nobody is watching?
The scanner you already pay for
Checks headers, versions, and configuration. Never attempts an exploit. When the reviewer asks "has anyone actually proven this?", the honest answer is no — and its output is a list of "possibles" your team stopped triaging months ago.
The pentest you can't schedule in time
Real expertise, weeks of lead time, a point-in-time snapshot that ages in days. And when a finding is disputed, the dispute is a clash of opinions: nothing re-runs, so nobody can settle it.
The "AI-powered" tool that cried wolf
"The model believes there could be an injection." Your team burns an afternoon refuting a ghost. Ask any security buyer what breaks trust in a tool fastest: findings that turn out not to be real. The second time it happens, the report stops being read.
A hypothesis becomes a finding only when the exploit works in our sandbox and reproduces — twice, with a clean negative control. There is no step in our pipeline through which an unproven claim reaches your report.
A report your reviewer can interrogate.
You are not buying a scan or a document to believe. You are buying artifacts a skeptical third party can check on their own.
The interactive report
One self-contained HTML file, no login. Every finding ships with its PoC as a curl your reviewer can copy, paste, and re-run; the raw request and response; and a fingerprint of the whole report. They don't take your word for it — or ours.
The verified-resistances attestation
A one-page certificate of what held: the run's attack counts by outcome (attacked-and-held, adjudicated out, inconclusive), with date and the artifact fingerprint — hypothesis-level detail lives in the report, plus a fingerprint anyone can check against the certificate itself — it is self-contained and travels with the evidence. What resisted goes on the record, not into a shrug. The sample run's certificate is published — sample-attestation.pdf.
The retest, included
After your fixes, we re-run every PoC — one full retest pass, included. Each finding closes with a status: FIXED, still-vulnerable, or contested with evidence on the record. The engagement ends in a state, not a document.
Verifiable artifacts
The sample's sha256 is printed on the page and checkable against the published run artifact; the benchmark metrics are signed and the public key is published. The signature file publishes with the next re-measurement; the signed payload is available on request. A live certificate-status portal is on the roadmap — not claimed until it ships.
A "we found nothing" without a certificate is just an opinion.
Under the hood Every candidate finding executes twice, passes a negative control — the same probe with the payload removed must find nothing — and survives our own refutation before you ever see it. How it works →
Email to report in four steps.
Scope — one page, by email
You email engagements@naxxan.ai with your domain. We reply with a one-page scope: the target, the non-destructive rules, and what we may create while testing (declared, itemized). It gets signed by someone with authority over the system — payment alone authorizes nothing.
The agent runs — minutes per surface, 3 informed rounds
An autonomous red-team agent maps the surface, writes a PoC for each hypothesis, and executes them from an isolated operator environment. Execution runs under a scope a human signs first; sensitive capabilities (process execution, credential use, live-target actions) require a human gate, and redirects, WebSockets and filesystem reads outside the engagement are blocked by control, not by promise. The AI does not authorize itself.
Adjudication — we attack our own findings before you see them
Every candidate faces the validation door — fired twice against a clean negative control; HIGH-severity candidates additionally face adversarial adjudication before they ship: is this a real auth bypass, or a public endpoint doing exactly what it was built to do? A reflection, or a real outbound request? Reproduction kills the hallucinated finding; adjudication kills the misclassified one. What survives both is what you get.
Delivery and retest — within 48h of the agreed start
The report and the attestation land in your inbox. After your fixes, the included retest re-runs every PoC and closes each finding as FIXED / still-vulnerable / contested — the retest must be requested within 30 days of delivery (same scope); a contested finding is then re-tested until it closes.
Why 48 hours, if a run takes minutes? Because the machine's speed is not the product — the accountability is. The window covers the human gate on agent actions, adjudication, and a reviewed report before it ships. You get the agent's speed with a human signature on how it was used.
We publish even what we didn't find.
CYBERGYM · SUITE V1 · MEASURED ON OUR PUBLIC BENCHMARK SUITE (JUICE SHOP — VULNERABLE-BY-DESIGN, PUBLIC: THE ONLY TARGET WE CAN NAME) ────────────────────────────────────────────────────────────── Detection 2 of 4 known bugs (50%) · 3 runs · 95% Wilson CI [15%, 85%] (n = ground truth) False positives 0 — 9 findings adjudicated, 3 runs Run time minutes per surface (anchor: 374.1 s, run 7c30ed04dc73) Signature ed25519 · public key
Honest note A small suite with a wide interval — we print both. We publish even what we didn't find: suite v1 contains four known bugs and we land two; the two we miss are named, not buried (a BOLA we haven't landed within the run's budget, and an SSRF where the target reflects our probe but never makes the outbound connection — and that second miss is our attribution rule working: we refuse to count our own echo as a hit). Every figure here is measured on our public benchmark suite and signed with ed25519. Two numbers, one target, two modes, so we reconcile them here: the containment benchmark lands 2 of 4; the full-format sample run (374 s, run 7c30ed04dc73, see the sample report) validated one CRITICAL (admin auth demonstrated) and one HIGH 7.5 (read evidenced, class ceiling disclosed); the other nine: 5 attacked-and-held, 2 adjudicated out, 2 inconclusive — same target, different modes, both say so. And when our own adjudication kills one of our numbers, we withdraw it and correct ourselves in public — the full correction record lives here. If a vendor's benchmark never moves, ask who grades it.
In one engagement we delivered a CRITICAL finding. Adjudication killed it after delivery: the behavior reproduced, but the class was wrong — a public endpoint behaving as designed, not an auth bypass. We corrected the report and refunded it under our correction policy — the full record is public — and built the class check that keeps that error from being born again. If your security provider has never been wrong, ask how they would know.
The sample report
Against OWASP Juice Shop (public, vulnerable by design): 11 hypotheses raised: 1 CRITICAL (admin authentication demonstrated) + 1 HIGH 7.5 (arbitrary read evidenced; 9.8 class ceiling disclosed), 5 attacked-and-held, 2 adjudicated out, 2 inconclusive — each finding executed twice with a negative control, a 374 s run, generated end-to-end by the agent with no manual correction of the content.
One authorized domain, verified live
Direct, and with nothing to install. The sequence is on the record, and it spans two engagements. The first raised 3 headline candidates and our own adjudication killed 2 of them — the CRITICAL, delivered and then killed as a public endpoint behaving as designed (report corrected, refunded under our correction policy), and an SSRF-HIGH that was our own reflection, not an outbound hit — leaving 1 validated finding delivered and 7 attack classes closed with verified-resistances certificate. In the engagement that followed, 2 more candidates died in our own adjudication before they ever reached a client. The full count: 5 headline candidates across two engagements, 4 killed by our own adjudication, 1 delivered.
The engineering floor
Underneath all of it: 1818 tests — 1730 portable + 29 Docker E2E + 59 Chromium E2E, same generated count everywhere, mypy strict, 0 errors. Counts are generated from the suite, not typed by hand. What cannot be tested cannot be promised.
// FIELD NOTES
One fixed price. No severity-contingent fee.
NAXXAN Verify External
USD 995 — the first five engagements signed and paid · USD 1,500 from the sixth
- One standard-scope target: one public app or API with its domain
- Validated findings with a reproducible PoC: zero "possibles"
- Deterministic CVSS v3.1 vectors; final severity adjudicated in class
- Stack-specific remediation guidance
- Post-fix retest included (request within 30 days): findings close as FIXED / still-vulnerable / contested
- Verified-resistances attestation with a verifiable fingerprint
- Non-destructive: no credentials, no repo access, nothing to install — any synthetic write declared
Full refund before the first authorized execution.
100% payment upfront (card or ACH). Payment does not authorize the test — the signed scope does.
What you pay for is the analysis, the adversarial validation, and retest accountability — not consultant time or infrastructure. No severity-contingent fee. The price is the price.
Adjacent — quoted separately
- Authenticated Deep-Dive
- Codebase Verification
- AI & MCP Red Team (model behavior, prompt injection, tool and agent actions — quoted separately)
Visible and requestable through the same asynchronous flow; price by proposal. Running agents in production? The third one is the conversation to have.
NAXXAN Continuous — where we're headed. Recurring verification that starts from everything the previous one learned. It is not operational yet, so we don't sell it: join the waitlist → engagements@naxxan.ai
// FAQ
01How long does it take, total?
48 hours from the agreed start (the signature date, unless a later start is agreed) to the report and attestation in your inbox. Most of your own time is reading one page and signing it. The agent's runs take minutes; the rest of the window is the human gate on agent actions, adjudication, and report review.
02Is it legal for you to run this against my production?
Nothing runs until a one-page scope is signed by someone with authority over the system — the signature authorizes; the payment does not. The verification is non-destructive against the one target you authorize: no destructive payloads, no real data exfiltrated, no denial of service — and nothing is modified. The synthetic writes we may need (a test account, a record) are declared and itemized in the scope before the run and in the report after it.
03What exactly is a "standard-scope target"?
One public application or API with its domain — what an outside attacker sees, verified as a conventional web/API surface. That is the unit Verify External is priced on. An application that uses AI does not lose its conventional routes: they are in scope; what moves to AI & MCP Red Team (quoted separately) is testing model behavior — prompt injection, tools, and agent actions. Additional targets, authenticated surfaces, and codebases are also adjacent services.
04Who is behind NAXXAN?
A security engineer who builds in the open — NAXXAN is led by its founder, an engineer resident in Florida (US), trading individually until NAXXAN LLC completes formation (the entity details land on this page the day the filing does). We don't run a padded team page: the verifiable part of who we are is on this page — the public benchmark, the signed metrics, the sample report, and the corrections we published when we got things wrong. One email, engagements@naxxan.ai, reaches the engineer doing the work.
05What if you find nothing?
You keep the verified-resistances attestation: the run's attack counts by outcome (attacked-and-held, adjudicated out, inconclusive), with date, the artifact fingerprint and a hypothesis ledger — one page. A "we found nothing" without a certificate is just an opinion.
06Do you write to my systems or touch my data?
Nothing destructive: the PoCs prove a fault exists without destroying anything, exfiltrating real data, or taking the service down — and nothing of yours is modified. The synthetic writes we may need (a test account, a record) are declared and itemized in the scope before the run and in the report after it.
07If you find something critical, do you charge more?
No. There is no severity-contingent fee — we do not charge more for finding something severe. That incentive inflates severities, and inflated severities are the fastest way to lose your blue team's trust. And when one of our own findings does not survive our refutation, we correct the report. It has already happened, and how we handled it is public.
08How do I verify your numbers are real?
Start with the sample report: full findings, no signup, PoCs you can re-run yourself today — its sha256 is printed on the page and checkable against the published run artifact. The benchmark metrics are digitally signed and the public key is published. Every finding in a real report can be re-executed on request.
09Do you need access to my code or credentials?
Not for Verify External. We verify your authorized domain from the outside — exactly what a real attacker sees. Codebase and authenticated verification are adjacent services with their own scope.
10Who executes — a human or AI?
An autonomous red-team agent runs the assessment; the sample report took 374 s end to end. Execution runs under a scope a human signs first, and agent actions during live operation pass a human gate.
11How do I start?
Email engagements@naxxan.ai with your domain and a time window. We reply with the one-page scope; you sign it; the verification runs; the report lands within 48 hours of the agreed start (the signature date, unless a later start is agreed). Calls are possible, never required.
12What about continuous verification?
Where does my data go? Target-derived content (surface data, findings, response excerpts) is transmitted over TLS to and processed by our model provider under its API terms — the current provider publishes that API content is not stored and is not used to improve its services without explicit agreement; we verify the terms per provider and model before any change and disclose the provider per engagement, before processing your data. We do not control provider-side compliance. On our side: your report and technical evidence are deleted 90 days after the retest completes — or, if no retest is requested, 90 days after the request window expires or the engagement closes, whichever comes first; the signed contract package is kept seven years for legal and financial proof. Early deletion on request is honored except for records we must keep under legal, contractual or tax obligations, or evidence under a documented dispute hold. Target-derived technical content is never sold, published, or used for any other purpose; administrative data (billing, signatures, email) is processed by the providers that perform those functions (payment, email, e-signature) under their terms. The current provider and model are disclosed in your engagement's run manifest.
Where we're headed, on a waitlist. We do not sell what is not operational yet.
One email starts it.
Email engagements@naxxan.ai with your domain and a time window. We send back the one-page scope — target, non-destructive rules, what we may create while testing — you sign it, the verification runs, and within 48 hours of the agreed start you have the report, the attestation, and a retest waiting for your fixes.
For Verify External we need nothing else from your side: no credentials, no repository, nothing to install. Prefer a call first? Possible. Never required.
suggested subject: Verify External — [your domain]
$995 for the first five engagements · full refund before the first authorized execution