AI-Augmented Penetration Testing: how we use AI agents without sacrificing human judgment

Artificial Intelligence is changing many aspects of our lives. Everyone is using it, including malicious actors who are leveraging its potential to accelerate attacks against IT systems, opening up new possibilities for cybersecurity professionals.

In this scenario, penetration testing must evolve.

AI-Augmented Penetration Testing integrates agentic AI tools into human-led activities to accelerate lower-value operational tasks and dedicate more time to analysis, developing attack scenarios, and validating vulnerabilities.

It is not automated penetration testing. It is “human-led”, “AI-accelerated” penetration testing.

In this article, we explain our approach and why we have chosen to use AI to amplify the analyst’s expertise, not replace it.

The pace of attacks has changed. Defenses need to keep up.

For decades, Offensive Security analysts have carried out fundamentally manual work: building attack chains, combining heterogeneous tools, and applying years of experience to find what others miss. That work has not disappeared — but the context around it has changed radically.

On the one hand, the attack surface has exploded: cloud, APIs, software supply chains, and now AI systems themselves: LLMs, RAG applications, autonomous agents are introducing classes of vulnerabilities that did not exist five years ago.
On the other hand, attackers have already adopted AI: large-scale reconnaissance, faster exploit weaponization, and social engineering made more convincing by deepfakes. The result is an imbalance: whilst an organisation releases dozens of deployments between one test and the next, the traditional penetration testing model struggles to keep up.

2026 is the turning point when this tension becomes impossible to ignore.

Two models, one clear choice

The market has responded by splitting into two different schools of thought. On one side are fully autonomous platforms, which promise to replace humans — the analysts — with agents that orchestrate the entire attack. On the other is the “expert-led”, AI-augmented approach, where AI acts as a force multiplier for the specialist.

This distinction is not merely academic. Autonomous agents excel in terms of scope and continuous coverage, but judging what constitutes a genuine risk, validating a finding, and taking responsibility before a regulator remain human tasks. A flood of false positives generated by a machine is not a penetration test: it is triage work offloaded onto the client.

For a company called Betrusted, the choice is a natural one. AI amplifies the analyst’s judgment; it does not replace it.

Where Betrusted stands

This approach is in our DNA. Our published research – from the analysis of CVEs in the Linux kernel, to red-teaming language models with evolutionary prompts, to relay chains targeting Active Directory Control Servers – is proof that we know where automation ends and expertise begins. We know how to write exploits by hand; that is why we know exactly which parts of the work to delegate to an agent and which not to.

We also know how to maintain the right distance: Betrusted tests, it does not remediate. This structural independence ensures that test results are never shaped by the need to sell remediation services.

AI-Augmented Penetration Testing is the first step in an AI-augmented direction for our entire offering. It is the starting point, not the destination.

How AI-Augmented penetration testing works

The operating principle is simple: human-led, AI-accelerated. The human expert leads the engagement from start to finish, while frontier agentic tools reduce the time spent on phases that have historically consumed significant portions of an analyst’s day.

• Reconnaissance and enumeration. AI correlates data across extensive attack surfaces and rapidly generates enumeration scripts tailored to the target, freeing the analyst to focus on interpreting the results.
• Custom tooling and harnesses. Parsers, protocol clients, fuzzing harnesses, glue code: the “plumbing” that typically takes hours and, due to budget constraints, often simply remains out of scope. AI produces it in a fraction of the time, and the analyst refines it.
• Analysis of large codebases. In white-box engagements, AI triages large codebases, flagging suspicious sinks and data flows; every signal undergoes manual verification before becoming a finding.
• Exploit development. The scaffolding of proofs of concept and exploit chains is accelerated, but weaponization and validation remain in the hands of the specialist: we report demonstrated vulnerabilities, not hypotheses.
• Reporting. Verified findings are turned into clear technical and executive narratives more quickly, without sacrificing clarity.

One non-negotiable rule applies throughout: every result is manually validated by an analyst, and the client’s code and data never leave the agreed perimeter. No sensitive information is sent to unauthorized public models. It is the only way to ensure that AI is compatible with our ISO 27001 certification, the requirements of DORA and NIS2, and the trust that our name represents.

The benefits of AI-Augmented Penetration Testing

From a technical perspective:

• More attack surface covered within the same time frame. Accelerating enumeration and tooling frees up time that can be reinvested in broader coverage and deeper analysis.
• Tooling that was previously out of reach. Custom harnesses and exploits that, due to time and cost constraints, would not have been developed in a traditional engagement.
• Deeper exploitation. Weaponized and verified PoCs, not just theoretical vulnerabilities.
• Shorter time to report. The organization receives actionable findings sooner.

From a business perspective:

• More value per engagement day: the same expertise delivers more.
• A level of depth previously reserved for premium budgets, now made more accessible.
• A reduced risk window: faster reporting means faster remediation and less exposure.
• Evidence to support compliance: demonstrated results, not assurances, to meet regulatory requirements.

The first step

Offensive Security will not go back to what it was yesterday. The question is no longer whether to integrate AI, but how to do so without losing the judgment that makes a penetration test reliable. Our answer is AI-Augmented Penetration Testing: agents for breadth and speed, experts for judgment, validation, and accountability.

This is the first service in an AI-augmented roadmap. If you want to understand what this means for your attack surface, contact us.