Articles
What the security audit covered, and what it did not
The method and findings of the two AI-assisted audit rounds of November 2026, what was fixed, and what is still unverified.
About 4 min read · Last updated
In this article
Before you put a panel in front of your customers' servers, you want to know what "it has had a security audit" means. For Lightning VM the phrase means something narrow, and the README does not hide that.
What was done
In November 2026 an AI-assisted security audit was carried out in two rounds. It is not an audit by an independent third party, and both rounds are partial.
- The first round looked at the whole repository: the API, the worker, the web interface, the console gateway, the adapters and the installer.
- The second round looked at the areas the first had not examined, and re-audited every fix of the first round.
The method was the same in both: source review, and limited trials with mock data in a sandbox without network access. No request was sent to any running system.
What was found
Across the two rounds 48 weaknesses were confirmed, and all of them were fixed.
| Round | Findings | Breakdown |
|---|---|---|
| First | 22 | 1 high, 10 medium, 11 low |
| Second | 26 | 10 medium, 15 low, 1 informational |
The first round found no way, in the customer routes, for one customer to reach another's server, invoice, ticket or network. The second round did find such ways, each depending on a precondition a customer cannot choose alone, and they were fixed: during the transfer of a server, and on Proxmox after a VM number had been reused.
Examples of what was fixed
- A return from Param no longer counts as proof of payment.
- The previous owner's keys and password do not remain on a transferred server.
- Directory and SSO accounts cannot sign in by other routes.
- Backups are signed with the key.
- An "on behalf of" session ends when the staff member's authority is withdrawn.
- After an old backup is restored, the panel does not act on old records; it starts on hold.
- The password-reset token is not written to the request log.
- Console tickets issued before a revocation are refused.
Did the first round's fixes hold?
Of the 22 fixes, 17 were re-audited on a real PostgreSQL and hold. The re-audit found five places where a fix was incomplete, and all five were corrected. Four (directory and SSO sign-in) could not be re-audited in the second round; they were looked at later by source review only.
Two follow-ups that are not audit rounds
On 18 November 2026 what the second round had left open was dealt with through source review and tests. On 19 November 2026 the 65 review units the first round had closed without a finding were read again, and four more holes were closed. The README gives the same warning for both: these are not independently confirmed findings and are not counted in the 48. The second reading was done by a single reviewer and by reading the source only; nothing was reproduced independently.
What was not looked at
The README names the places that were not re-read line by line:
- the two rounds of fixes in the billing code;
- the identity, support, notification and chat modules;
- the installer's backup, update, domain, host and uninstall paths;
- the adapters other than Proxmox;
- the clearing of browser storage and the saved-views component.
What was not verified
The fixes were verified with unit tests and with integration tests on a real PostgreSQL and Redis. They were not verified with:
- real hypervisors;
- a real Proxmox cluster;
- a real Param merchant account;
- a real directory server and SSO provider.
The installer's paths were tried with mock files only.
What this means for you
- Try the panel in your own environment and take backups before you expose it to the internet.
- Make two-factor authentication mandatory for admins; Insights show admins who have it switched off.
- If you are going to use LDAP or SSO, try the sign-in and permission flows from start to finish with your own provider; this path was not verified with a real one.
- If you have someone who can read the code, have them read it. The project was written largely with AI and, as the README puts it, contains code that no human has reviewed.
If you find a vulnerability, do not open a public issue; follow the route in SECURITY.md in the repository (on GitHub: Security → Report a vulnerability).
Sources: docs/security-audit.md (English), and the "Güvenlik" and "Bilinen sınırlar" sections of the README (Turkish).
