Aucun article n'a été trouvé.
Aucun article n'a été trouvé.
Pentesting
-
4
min de lecture
-
November 17, 2026

What Is Agentic Pentesting? A Complete Guide

-
- -
What Is Agentic Pentesting? A Complete Guide

Security teams are being asked to assess applications that change far more quickly than a conventional penetration-testing program can accommodate. New code, cloud services, APIs, and identities expand the environment between engagements, while the work required to scope and authorize a test can take six to twelve weeks. A rigorous assessment can therefore cover a version of the application that has already moved on.

Agentic pentesting applies AI agents to the work between an agreed scope and a verified finding. The system investigates an application, forms and tests hypotheses, uses security tools, and adapts its next steps to the evidence it receives. Buyers should assess whether that complete system can conduct an authorized test at sufficient depth without introducing a new operational risk.

What is agentic pentesting?

Agentic pentesting is penetration testing performed by software agents that can plan, reason, act, observe results, and revise their approach within defined boundaries. A scanner looks for known patterns, while scripted automation follows a prewritten sequence. An agentic system can pursue a goal through multiple steps and choose tests in response to what it discovers about the target.

An agent might map an application’s routes and technologies, identify an authenticated API that handles invoices, and investigate whether the authorization checks consistently separate one tenant from another. It may establish a usable session, find the object identifiers the application expects, submit a bounded request, and determine whether the returned data proves cross-tenant access. The model provides part of the reasoning ability; the system also needs target context, purpose-built tools, memory, coordination, and an evidence standard. A frontier AI model is not a pentest on its own.

How agentic pentesting works

The workflow follows the broad logic of an experienced tester, even when parts of it run concurrently. It begins with reconnaissance: mapping pages, endpoints, technologies, and behavior, then combining this information with the agreed scope and supplied credentials. Continuous attack surface management improves the starting point by maintaining a view of exposed assets, technology changes, and known exposures.

The system develops hypotheses that are specific enough to test, perhaps asking whether an API enforces a role boundary or whether an input reaches a sensitive database query. Specialist agents can investigate different exposure classes in parallel, while an orchestrator tracks the overall investigation and redirects effort when a result opens a credible next path.

Every action remains bounded by the test’s permissions. The system collects evidence, compares it with the hypothesis, and discards the lead, strengthens the case, or takes the next permitted step. A credible result demonstrates a failed security property and records how it was demonstrated. This is the standard behind proof-backed findings, which avoids handing developers an issue they have to investigate from scratch.

How is agentic pentesting different from AI-assisted testing?

AI-assisted testing may refer to a copilot that helps write a script, a model that summarizes scan output, or a tool that suggests payloads. These functions leave the main work of choosing a path, running it safely, interpreting the result, and deciding the next step with a human operator.

Agentic pentesting connects these activities into a supervised loop. The system maintains a goal, chooses a next action from a constrained set, gathers new information, and uses that information to continue the investigation. Its autonomy is bounded by technical controls that limit targets, methods, rate, credentials, and the effects an agent can cause. Prompts cannot enforce these limits by themselves.

Where agentic pentesting fits alongside traditional pentesting

Agentic pentesting changes the cadence of testing, particularly for internet-facing applications and APIs that change frequently. It can be used before a release, after a major infrastructure change, when a new asset appears, or as coverage between scheduled engagements. Teams can investigate more of the work that would otherwise wait for the next annual or semiannual test. Continuous penetration testing explains why this shift matters as environments change.

Traditional human-led testing remains important for a deep review of a business process, formal assurance, and exploratory work outside the platform’s defined test categories. Human judgment is also needed when a technically valid finding has to be interpreted in the context of a business and its risk appetite. A layered program combines continuous asset context, agentic testing, and human specialist review where it is most valuable. Our guide to what to look for in AI-powered pentesting explores this division in more detail.

What makes agentic pentesting safe enough for production?

Safety is a property of the system surrounding the model. Buyers should understand what the platform guarantees when a model makes a poor decision, encounters malicious instructions in a response, or continues an investigation too aggressively.

Ask how scope is enforced, including whether an out-of-scope request can physically leave the platform. Ask which tools and methods agents can use, which actions are prohibited, and how the system prevents production-impacting activity such as deleting assets, creating charges, or changing account settings. Rate limits should be shared across concurrent agents so a broad investigation does not become an unpredictable load on the target.

The platform should also protect credentials, log actions, check for prompt injection, and stop work on request. A complete action trail and clear test window give security operations teams the information they need to distinguish approved testing from an incident.

What are the limitations of agentic pentesting?

Agentic systems can investigate quickly and persistently, yet their effectiveness depends on the environment they are given. Authentication remains a recurring challenge because every application carries identity and session state differently. A browser may be logged in while a separate testing tool lacks the cookies, headers, or tokens required to make an authenticated request. The system has to recognize and correct that problem before its results can be trusted.

Models are creative within the categories, tools, and permissions the system has defined. That boundary supports safety and repeatability, although a platform will not discover every novel business-logic flaw that a senior human tester could uncover through a long exploratory engagement. Security and engineering leaders also retain the decision about whether a technically valid finding matters to their business. Someone must set the scope, receive the results, and make sure remediation happens.

How to evaluate an agentic pentesting vendor

When evaluating a vendor, ask for a demonstration of a full test journey instead of a list of model capabilities. The questions below reveal more than a claim that a product uses agents.

  • Can the system show its scope, action log, and evidence behind every reported finding?
  • How are target restrictions, prohibited actions, and rate limits enforced at the tool and network layers?
  • How does it establish and validate authentication across browsers, APIs, and other tools?
  • What separates an unverified signal from a finding, and who reviews the result before it reaches engineering?
  • How does the platform coordinate concurrent agents and retain context across a test?

The answers should make clear whether the product offers an investigation system with governed autonomy. For a broader view of the market, see The AI hacking boom: seventy tools in eighteen months.

{{related-article}}

What Is Agentic Pentesting? A Complete Guide

{{quote-1}}

“
”
,

{{quote-2}}

“
”
,
obtenez une démo de 15 min

Commencez votre voyage dès aujourd'hui

La plateforme de sécurité offensive de bout en bout de Hadrian se déploie en quelques minutes, fonctionne de manière autonome et fournit des informations faciles à exploiter.

Ce que vous allez découvrir

  • Surveiller les actifs et les changements de configuration

  • Comprendre le contexte des actifs

  • Identifier les risques, réduire les faux positifs

  • Prioriser les risques à fort impact

  • Simplifier la remédiation

The Hadrian platform displayed on a tablet.
No items found.