Demonstrated
The agents exploited the flaw within the scope. A hypothesis or a theoretical risk is not published.
Methodology
Eight steps, from declaration to retest.
Nothing is tested until you have proved control of the domain and approved the scope. Here is what happens at each step, what you provide, and where the test stops.
Walkthrough
What happens at each step.
Declaring the application
You create your space and declare the application’s exact URL. It is recorded as a target: nothing is tested at this stage.
What you provide
The application’s HTTPS URL.
Domain verification and authorization
You publish a single DNS TXT record on the domain; SpectrAI checks it. By accepting the quote and the terms, you explicitly authorize testing this scope.
What you provide
Access to the domain’s DNS zone, and the decision to authorize the test.
Defining the scope
You list the URLs or journeys to test and the excluded paths. You describe the context, the priorities and the validation rules: this is the guidance given to the agents. SpectrAI checks that the application fits the standard scope.
What you provide
The URLs, the exclusions, the business context and, if you have one, the API documentation (OpenAPI, Swagger or Postman export).
Accounts and constraints
You provide the test accounts, one per role, and your constraints: paths not to touch, maximum request rate, careful mode for production, start window. The accounts are stored encrypted.
What you provide
Up to two dedicated test accounts, with no real personal data, and your operating constraints.
Controlled execution
AI agents attack the application within the scope, with a budget and a turn cap set before launch. You follow their requests and their tools live, and you can stop the test at any time.
What you provide
Nothing, unless you want to follow or stop the test.
Analysis and human review
Only the vulnerabilities of the last successful run, with consistent proof, are put forward for review. A human reviews each one: they keep it or reject it before publication. Severity is recalculated from the CVSS vector.
What you provide
Nothing. You receive the outcome of the review with the report.
Report delivery
The report is frozen: scope, coverage declared by the agents, vulnerabilities with their proof, impact, CVSS, CWE and recommendations, attack chains when they are demonstrated. It exports as PDF, JSON, SARIF, CSV and Markdown.
What you provide
Nothing. A result with no demonstrated vulnerability is a valid result and also produces a report.
Retest
Once a vulnerability is fixed, you mark it as fixed and request the retest. A targeted test replays the original proof and its obvious bypasses, within 60 days of the report. Verdict per vulnerability: fixed or still present.
What you provide
The deployed fix and the retest request. In the application, the action is called “Verify the fix”.
Vulnerability
What counts as a vulnerability.
An alert is not a vulnerability. A published vulnerability is demonstrated, proved and reviewed.
Proved
The proof (requests, responses, reproduction steps) is kept with the vulnerability. Cookies and secrets that are quoted are masked.
Reviewed by a human
A human reviews every vulnerability before publication. Without consistent proof, it does not enter the report.
Qualified
Each vulnerability is published with its impact, CVSS severity, CWE weakness and a fix recommendation.
What is not published
- A risk score with no exploitation.
- An old software version, or a dependency CVE whose exploitation is not demonstrated.
- A finding with no consistent proof.
Limits
Where the test stops.
What SpectrAI does not do, and what it does not guarantee.
External web only
Web applications reachable over HTTPS from the Internet, under a domain you proved. No internal infrastructure, no mobile app, no source code.
No guarantee of discovery
A pentest does not prove the absence of vulnerabilities. The coverage shown in the report is declared by the agents, not measured.
Neither PASSI qualification nor guaranteed compliance
SpectrAI is not a PASSI-qualified provider. The report provides dated evidence an auditor can use; it guarantees no compliance (ISO 27001, PCI DSS, NIS2, DORA).
Controlled exploitation, not guaranteed
The agents receive mandatory instructions: no denial of service, no destructive testing, no persistence, no pivoting. These are instructions given to AI agents, not a technical guarantee. Only the start window is enforced by the platform. For a critical application, test a pre-production environment where you can.
Outbound traffic not isolated
The agents’ outbound traffic is not yet isolated at network level: keep the test to domains you own.
Third-party AI model
A third-party language model analyzes the application’s responses and your instructions. No processing within the European Union is guaranteed at this time. The privacy policy names the provider.
Test accounts in clear text during execution
Test credentials are stored encrypted, but the agents use them in clear text during the test. Provide dedicated accounts, with no real personal data, and revoke them afterwards.
Go further
Read a report before you order.
The sample report shows what you receive: scope, demonstrated vulnerabilities, proof, impact, recommendations and limits.