AI-Assisted Penetration Testing

Artificial intelligence is introducing new capabilities into offensive security. While conventional security tooling remains valuable for identifying potential vulnerabilities, autonomous AI agents can take technical investigations further by analysing findings, developing attack paths and attempting exploitation.

At DeepBlue, we have developed a proprietary AI-assisted penetration testing framework that incorporates our experience in penetration testing, red teaming and exploitation techniques. By deploying multiple autonomous agents within a single assessment, we can investigate different attack surfaces in parallel, accelerate technical analysis and streamline the overall testing process. The objective is not to replace our penetration testers, but to extend their capabilities while maintaining direct human oversight and strict operational control.

Extending our manual-first approach

DeepBlue has always approached penetration testing as a predominantly manual discipline. Automated tools help us map the attack surface and identify potential weaknesses, but their output primarily serves as the starting point for our specialists. The actual penetration test involves interpreting these results, investigating application logic, developing attack paths and demonstrating whether vulnerabilities can be exploited.

Our AI framework builds on this methodology. Autonomous agents can perform reconnaissance, operate existing security tooling and independently investigate the resulting technical information. Rather than simply reproducing automated scans, they formulate hypotheses, develop testing scripts and adapt their approach based on their findings. This allows our operators to delegate substantial parts of the technical investigation while concentrating on complex exploitation scenarios and vulnerabilities that require specialist judgement.

Translating offensive security expertise into autonomous agents

The framework uses a multi-agent architecture that distributes technical investigations across specialised agents within the same customer engagement. An initial reconnaissance phase identifies relevant systems, application components and potential attack surfaces. Based on these results, multiple agents investigate different areas of the same environment simultaneously, such as authentication, authorization, cross-tenant isolation, API security and application logic.

These agents operate according to methodologies developed from our practical experience in penetration testing and adversary simulation. We have incorporated our established testing procedures, exploitation techniques and methods of investigating attack paths into their workflows. This enables agents to approach technical challenges using the same principles that guide our specialists, rather than relying exclusively on generic AI reasoning or predefined vulnerability checks.

For example, when investigating a web application, an agent may independently download and analyse JavaScript bundles, identify undocumented API endpoints and develop scripts to interact with them. It can subsequently investigate authorization mechanisms, manipulate requests and explore potential attack paths. When an approach fails, the agent can reconsider its findings and pursue alternative techniques.

From suspected vulnerabilities to verified exploitation

A potential vulnerability is not necessarily exploitable. This distinction is particularly important when working with AI, as language models can produce technically plausible conclusions that do not reflect actual weaknesses in the target environment.

Our framework therefore focuses on practical validation. Agents attempt to develop working proofs of concept (PoCs) that demonstrate identified vulnerabilities. For example, when investigating a multi-tenant application, an agent might examine object references and authorization controls to determine whether information belonging to another tenant can be accessed.

Successful exploitation attempts produce technical evidence and reproducible PoCs for our operators to review. When an agent cannot demonstrate exploitability, the suspected vulnerability is rejected or recorded as an informational observation, depending on the available evidence. Our specialists subsequently validate the findings, investigate potential attack chains and determine the actual security impact.

Autonomous execution with human oversight

Autonomous exploitation introduces operational risks that require more than instructions telling an AI model what it should or should not do. DeepBlue has therefore developed a dedicated execution framework that enforces technical boundaries independently of the underlying AI models.

Within these boundaries, agents can autonomously execute commands, interact with applications, develop scripts and perform exploitation attempts. The framework restricts network interactions to explicitly authorized targets and controls which tools and capabilities agents can access. Testing is typically conducted against dedicated development, test or acceptance environments to minimise operational risks.

For activities that could affect system integrity, availability or confidentiality, the framework implements mandatory human approval gates. State-changing operations, data exfiltration and other predefined high-risk activities require explicit authorization from an operator before execution. These restrictions are technically enforced, preventing agents from independently proceeding with operations that require human intervention.

Our operators actively supervise the assessment, review agent activity and determine when additional investigation or changes to the operational approach are necessary. This human-in-the-loop model combines autonomous technical execution with the operational discipline and accountability required for professional penetration testing.

Data protection and model deployment

Penetration testing involves sensitive technical information, including infrastructure configurations, application vulnerabilities and authentication mechanisms. Protecting this information remains essential when integrating AI into our testing methodology.

DeepBlue uses cloud-hosted AI models deployed within the European Union, under enterprise arrangements that protect customer information and prevent its use for model training. The models provide analytical and reasoning capabilities, while our proprietary framework manages tool execution, scope enforcement and operational permissions. This architecture allows us to use advanced AI capabilities while maintaining defined controls over offensive activities and customer data.

Keeping the penetration tester in command

AI-assisted penetration testing changes how technical investigations are performed, but not who is responsible for the outcome. Our senior penetration testers remain in control of the assessment, establish its objectives, supervise autonomous activities and validate the resulting findings.

By delegating reconnaissance, repetitive technical analysis and parts of the exploitation process to specialised agents, we can investigate different attack surfaces within an engagement simultaneously and reduce the time required for certain testing activities. This allows our specialists to focus on complex attack paths, business logic vulnerabilities and scenarios that demand creativity, experience and contextual understanding.

The result is a more efficient penetration testing process that combines autonomous technical investigation with the expertise and quality assurance of experienced offensive security specialists.