AI made finding weak spots fast and easy. For attackers too.
Assume your website and your code will be read by AI, by someone who is not on your side. Have it read first, by people who are.
For agencies, programmers and hosts
An extra service for the clients whose websites or software you build, maintain or host. Six questions, answered.
What exactly do I offer?
A security scan of your client's website or code base by Claude Mythos 5.2, and a report in which each finding comes with its evidence, its severity and a suggested fix (Section 6). The code need not belong to a website: an app, a back-end service or an internal tool is read the same way. It is not available off the shelf: Anthropic releases Mythos 5.2 only to organisations it has verified for security work, 2 and we are one of them.
Who sees the report?
You do, first. An agency or a programmer sees what was found in the sites and the code it looks after before its client does, and decides how to present it. A host passes the report on to its customer, or has us send it. What we typically find (Section 2) sits in plugins from third parties.
How much work is it?
Arranging a scan is little work: you choose the client, get their signature on our permission form and give us read-only access. The report follows within [working days]. Fixing what it finds is separate work, and it is not ours: the report suggests a fix for each finding. An agency or a programmer does the fixes itself; a host's customer has them done by their own developer.
How do I earn on it?
You pay us a fixed rate for each scan and set your own price to your client. We give you our rate on the first call; we do not publish prices. For an agency or a programmer, the fixes are a second source of paid work.
Is it right for us?
It suits agencies and programmers whose clients run online shops, apps or other software their business depends on, and hosts with managed business customers. It is not a few-euro add-on for thousands of sites: each scan is scoped and supervised by a person.
Who is Neumann Labs?
A diagnostic laboratory in Budapest that came to security through its own software (Section 8). We are verified under Anthropic's Cyber Verification Program. 4
Start with one client
Pick one client site or one repository. We scan it, and you see the report before you decide anything further. [terms of the first scan]
Talk to us By call or by e-mail. Tell us roughly how many client sites or code bases you look after.
When to have a scan, and why
Two reasons apply to almost every company, whatever the moment.
A reputable source is no guarantee.
In our experience, even plugins from reputable suppliers can turn out to be unsafe.
It has never been scanned like this.
Whatever checks your website and your code have had, none read them with the depth and power described in section 5.
Beyond that, a scan is worth most when your systems change hands or change shape. Six occasions call for one.
- Before you launch
- A new website, shop or app. A weak spot is far easier to fix before customers depend on the thing than after.
- After a big change
- A new payment step, a new login, a rebuilt part. New code is where new weak spots arrive.
- When you inherit code
- An agency hands over, a developer leaves, you buy a company. It is yours now. Find out what is in it.
- When code has grown old
- Nobody has opened it in years, but it still runs and it can still be reached.
- When someone asks
- A customer, an insurer or an auditor wants to know what you have checked, and when. A dated report answers them.
- At regular intervals
- Your code keeps changing, and so do the tools used against it. What was clean last year may not be now.
What we typically find
Four findings from our own scans of online shops, in plain words.
One web request can switch off card payments for the whole shop.
No login needed.
A fake “payment received” message is enough to get the goods.
The shop does not check who sent it.
A shopper can check out with the VAT removed, without a valid VAT number.
The shop owes the tax anyway.
The file that holds the shop’s bank key can be downloaded like a picture.
How the scanning performed in an early trial
Before offering the scan to others, we ran an early trial on the plugins that online shops are built from. Table 1 is the record of that trial: what was scanned, what was found, and what held when we put it to the test.
| Count | What it counts | How to read it |
|---|---|---|
| 85 | Plugins and code libraries scanned, about. | Most are payment, shipping and invoicing plugins for online shops. |
| 150 | Findings recorded across all scans, about. | Every weak spot the scans reported, from minor to serious. Most have not been tested. |
| 28 | Findings put to the test on our own test shops. | The ones we judged most serious. |
| 27 | Of those 28, confirmed as real flaws. | In 16 plugins and libraries. Five are rated high, 21 medium, one is not yet rated. |
Twenty-seven of the 28 findings we tested were real flaws. In the one that was not, the plugin did what the scan said, but WordPress itself blocked the result.
What testing corrected more often was the detail: the conditions a flaw needs, and how serious it is. That is why we offer to confirm critical and high findings ourselves, as an option (section 6).
Request a scan Find out whether any of these is true of your systems.
What we scan
Your code, your website, or both. The code need not belong to a website: an app, a back-end service, an internal tool or a library is read the same way. A code scan reads every file in scope from the inside (Fig. 1). A website scan maps and scans what is public from the outside and follows each process, such as login or checkout, step by step. Ordered together, what we see from the outside is traced back to the line of code that causes it, so your developers know exactly what to change.
- app/your own coderead
- config/settings, and secrets left in coderead
- auth/who may sign inread
- api/read
- billing/read
- jobs/read
- legacy/untouched for yearsread
- infra/deployment filesread
- vendor/dependencies and their known weak spotsread
We need: read-only access to the repository, agreed in writing. Request a code scan
- /home pagescanned
- /loginlogin formscanned
- /checkoutpayment stepscanned
- /accountcustomer accountsscanned
- /contactcontact formscanned
- /searchsearch boxscanned
- scriptsyours and third-party add-onsscanned
- /apiwhat the pages callscanned
- servercertificate and headersscanned
We need: written permission for the domains in scope. Request a website scan
- /loginauth/traced
- /checkoutbilling/traced
- /accountapp/traced
- /contactapp/forms/traced
- /searchapi/search/traced
- /apiapi/traced
- scriptsvendor/traced
- /adminadmin/traced
- serverinfra/traced
We need: both of the above. Request both
How a scan works
Agree the scope
What is in, what is out, when, and the fixed price. You sign written permission. For code, you grant read-only access.
Mythos 5.2 scans
The whole website, the whole code base, within the scope we agreed. We map, scan and follow each process. We do not attack, and we change nothing.
The findings are written up
Each finding is set out with its evidence, its severity and a suggested fix, and the list is ranked by risk. In the base scan we do not confirm the findings one by one: the evidence is there so that your developers can check each of them.
You get the report
Private, in our client portal, in plain words, with a suggested fix for each finding that your developers can work through. As an option, we confirm the critical and high findings ourselves (section 6). Ask us to scan again once the fixes are in.
All of it runs through our client portal. You sign in by invitation at [portal address]. There you hand over the code, see at every step where the scan stands, read and download the report, and write to our team. Every sign-in, view and download is logged, by your side and ours, and you can read that log yourself.
Why a traditional scanner is not enough
Scanners match patterns. Claude Mythos 5.2 reads. A scanner checks each line against a list of known mistakes. Mythos 5.2 reads a website or a code base as one system, and it can keep a large code base in view at once. Table 2 is how we see the difference.
| Pattern scanners | Our scans, with Claude Mythos 5.2 | |
|---|---|---|
| How it reads | Line by line, against a list of known bad patterns | The whole code base as one system: where data goes and who may do what |
| What it finds | Known patterns | Known patterns, plus flaws in logic and permissions |
| Large, old code | Same rules, whatever the size | Keeps the full picture, old folders included |
| What you receive | A long list of possible issues | Findings ranked by risk, each with its evidence and a suggested fix |
Not every AI will do, either. Many AI tools turn out security reports that read well and are wrong: flaws that do not exist, seriousness blown up. Developers call it AI slop. In our experience Mythos 5.2 is far more precise (Table 3, last row). No model is right every time. That is why each finding in our report comes with its evidence, so that a developer can check it, and why you can have us confirm the critical and high ones (section 6).
Anthropic describes Claude Mythos 5.2 as its most capable model for cybersecurity engineering. 1 What it has published about the model's performance is in Fig. 2 and Table 3.
| In short | Statement | Who says it | Source |
|---|---|---|---|
| 1 in 6 | One in six confirmed bugs found by Mythos 5.2 were missed by every other frontier model tested. | One security company, as reported by Anthropic | 1 |
| ≈ 2× | Roughly twice the vulnerability discovery results of Mythos 5.1, across Mozilla's scanning campaigns. | Mozilla | 1 |
| 1M | A context window of one million tokens: room to keep a large code base in view at once. | Anthropic's documentation | 2 |
| 0 | Findings whose seriousness Mythos 5.2 overstated in one tester's audit. Earlier models overstated about one in twenty. | One tester, as reported by Anthropic | 1 |
Standard scanners, alongside (optional). If you want both, we also run the established scanners: known vulnerabilities in your dependencies, secrets left in code, and certificate and header set-up. On websites we use only the kind that looks, never the kind that attacks. Their results go into the same report. Add it to a scan
What you get
One private report (Fig. 3). The owner can decide from it and the developers can act on it.
- A ranked list, most serious first
- For each finding: where it is, the evidence, why it matters, and a suggested fix
- A one-page summary in plain words for the owner
-
Critical
Confirmed
- Where
- Why it matters
- How to fix it
- HighConfirmed
- Medium
- Low
Option: we confirm the critical and high findings. In the base scan the findings are not confirmed by us one by one; each carries its evidence so that a developer can check it. If you want more certainty before you act, we test every finding rated critical or high ourselves, and mark it confirmed or tell you that it did not hold. We suggest fixes; we do not make them.
How we handle your code
You are handing over the thing your business runs on. Table 4 sets out what happens to it.
| Access | Read-only, agreed in writing, removed [when]. |
|---|---|
| Your customers' data | We do not need it and we do not ask for it. Where a login is needed, we use test accounts you set up. If we come across personal data, left in code or exposed by a flaw, we do not copy it. We stop there and tell you at once. |
| Where it is read | In [our environment], and by Claude Mythos 5.2 through Anthropic's service. |
| Safety monitoring | By default, using Mythos models requires data to be kept for 30 days for safety monitoring. 3 |
| Kept by us | Our copy is deleted [retention period] after the report. |
| Who sees it | [named reviewers], under a confidentiality agreement [to confirm]. |
| Shared with | Nobody else. Findings go to you only. |
| Where we work with you | In our client portal, by invitation only. Every sign-in, view and download is logged, and you can read the log (section 4). |
Who we are
Neumann Labs is a diagnostic laboratory in Budapest, run by Spring Molecular LLC. Running a laboratory led us to build our own apps, and protecting patients' data in those apps led us into cybersecurity. We now do for others what we first had to do for ourselves.
Talk to us
Tell us about your clients, or about the site or the code you want scanned. We reply within [response time]. Nothing starts before you agree.
References
- Anthropic, Introducing Claude Fable 5.2 and Claude Mythos 5.2, 9 November 2026. anthropic.com/claude-fable-and-mythos-5-2
- Anthropic, Claude Mythos 5.2, Claude Platform documentation. platform.claude.com/docs/en/models/mythos-5-2/overview
- Anthropic, Claude Mythos 5.2, product page, on data retention. anthropic.com/claude/mythos
- Anthropic, Cyber Verification Program, support article. support.claude.com/en/articles/14604842
Claude Mythos 5.2 is an AI model made by Anthropic. Neumann Labs uses it as a tool. Neumann Labs is not affiliated with, or endorsed by, Anthropic.
Neumann Labs · Spring Molecular LLC, Budapest · © 2026