Sovra Recon·Guide · AI pentest

AI Pentest and Agentic Pentest: The Complete Guide

By Solan Barlet, co-founder of Sovra Recon
Updated October 29, 2026

In short

An AI pentest, or agentic pentest, is a penetration test conducted by artificial intelligence agents capable of reasoning, chaining together actions, and adapting their attack based on what they discover—just as a human penetration tester would. Unlike a scanner, which compares a target to a list of known signatures, an agent explores the application, formulates hypotheses, and actually attempts to exploit them. When properly supervised, it produces proven vulnerabilities, along with proof of exploitation, in a matter of hours rather than weeks.

What is an agentic pentest?

A penetration test (pentest) involves attacking a system with the owner’s permission to uncover vulnerabilities that an attacker could exploit. Historically, two approaches have coexisted: the manual pentest, performed by experts—which is precise but time-consuming and rarely conducted—and the automated scan, which is fast but limited to known vulnerabilities.

Agentic pentests add a third approach. An AI agent is a program based on a large language model that is given an objective, chooses its own tools (browser, HTTP requests, scripts), observes the result, and decides on the next step. Multiple agents can divide up the work: one maps the system, another tests authentication, and a third attempts to bypass access controls.

The term “agent-based” refers to this autonomy: the agent does not follow a fixed script; it pursues a goal. This is what enables it to find business logic vulnerabilities that no signature can describe.

AI pentest, vulnerability scanners, manual penetration testing: what are the differences?

Vulnerability scannerManual penetration testAI pentest (agentic)
MethodCompares the target to known signaturesAn expert explores and attacksAgents explore, reason, and attack
Business logic vulnerabilities (IDOR, access control)RarelyYesYes
Proof of exploitationNo (presumption)YesYes, if the agent requires it to validate
False positivesManyRareRare if each vulnerability is replayed
DurationMinutesDays to weeksHours
Realistic FrequencyContinueAnnualOne-time or continuous

In practice, the quality of an AI pentest depends on the design of its agents: how they divide up the work, how they sequence the steps, and how they prove their claims. A well-designed agent only retains what it has actually exploited.

How does an AI pentest work?

The steps remain those of a traditional penetration test; what changes is who performs them and how quickly.

  • Scope and authorization: scope (domains, applications, APIs), time windows for intervention, rules of engagement, signed mandate. Without written authorization, a penetration test constitutes a criminal offense (Articles 323-1 et seq. of the French Criminal Code).
  • Reconnaissance: Agents map the attack surface—subdomains, pages, forms, API endpoints, and technologies in use.
  • Vulnerability research: injection, access control, authentication, session management, configuration, business logic.
  • Controlled exploitation: Each hypothesis is actually tested, without any destructive actions. A vulnerability that cannot be demonstrated is not retained.
  • Validation: The vulnerability is re-executed and verified before being included in the report, which eliminates false positives.
  • Report: For each vulnerability, the severity (CVSS score), proof of exploitation, impact, and recommended fix are provided, along with a summary that is easy for management to understand.

What AI Really Changes

  • Frequency: A test that takes hours can be rerun after every production deployment, rather than just once a year.
  • Coverage: Dozens of agents work in parallel across the entire surface area, whereas a human must prioritize.
  • The proof: by requiring actual exploitation, we replace lists of “potential vulnerabilities” with demonstrated vulnerabilities.
  • Accessibility: SMEs that have never conducted a penetration test can now assess their actual vulnerability.

Limitations to Be Aware Of

An AI pentest isn’t magic, and an honest service provider will tell you that.

  • Scope: An application test does not cover phishing, physical security, or your suppliers’ supply chain.
  • False positives: A language model may flag a vulnerability that does not exist. Only validation through actual exploitation, with supporting evidence, protects against this risk.
  • Security of the test itself: autonomous agents must be restricted (no denial-of-service attacks, no data tampering, and the ability to stop the test at any time).
  • Data: Verify where the models are running and what happens to the data collected during the audit.

AI pentest and Compliance (NIS2, ISO 27001, Insurers)

An AI pentest report is as valid as a manual pentest report if it contains the same elements: scope, date, method, proven vulnerabilities with their severity, recommendations, and verification of fixes. It then serves as evidence for Article 21 of NIS2 (assessment of the effectiveness of measures), for controls 8.8 and 8.29 of ISO 27001, or for a cyber insurance questionnaire.

Its unique advantage lies in frequency: whereas an annual audit provides a snapshot of a system’s state at a single point in time, frequent testing reveals an ongoing process—something that recent standards emphasize.

How do you choose an AI pentest provider?

  • Does it require a written mandate and a defined scope before any attack?
  • Is each vulnerability in the report accompanied by reproducible proof of exploitation code?
  • Is every vulnerability in the report proven by an actual exploit, or is it merely assumed?
  • What guarantees are there that the tests are non-destructive?
  • Where is the audit data hosted and encrypted, and how long is it retained?
  • Is the report actionable for your compliance requirements (ISO 27001, NIS2, insurer, clients)?

This is the approach we take at Sovra Recon: fully automated offensive AI agents that retain only what they have proven.

Frequently asked questions

Will AI pentest replace human penetration testers?

It is transforming their profession. Agents conduct exploration and exploitation on a large scale, continuously, and at a cost that makes testing accessible to all companies; penetration testers retain the tasks that fall outside the scope of automated testing, such as social engineering or physical intrusion.

Is an AI pentest legal?

Yes, under the same conditions as a manual penetration test: only on systems you own or for which you have written authorization, within a defined scope.

What is the difference between AI pentest and PTaaS?

PTaaS (Pentest as a Service) refers to a delivery model: a subscription with a monitoring platform. AI pentest refers to a method: the attack is carried out by agents. A service offering may combine both, such as a continuous agentic pentest.

How long does an AI pentest take?

For a typical web application, this takes a few hours, depending on the attack surface, the complexity of authentication, and the presence of protections such as a WAF.

Sources

Find out where an attacker would get in.

Our AI agents actually attack your scope, fully automated, and hand you the proof of every vulnerability.