Skip to main content

What if a pentest no longer ended with a list of vulnerabilities to fix, but instead enabled developers to address them while testing was still underway? Together with leboncoin, Yogosha explored an augmented white-box pentesting approach combining AI agents, human expertise and real-time remediation.

From Five Years of Bug Bounty to Augmented Pentesting

The partnership between leboncoin, part of the Adevinta Group, and Yogosha began in 2020 with a Bug Bounty program.

Since then, their approach to offensive security has evolved. After several years of continuous offensive security through Bug Bounty and a traditional pentest conducted in 2026, leboncoin wanted to explore a new approach across four applications: an Augmented Pentest conducted in a white-box environment.

This time, Yogosha’s experts were not limited to what an attacker could observe from the outside.

The leboncoin teams granted them access to the source code of 4 applications across 20 repositories, covering back-office systems, public-facing websites, APIs, shared libraries and infrastructure.

The objective was to combine the broad coverage provided by AI agents with the analytical capabilities of security experts to examine the code in depth, identify vulnerabilities and enable developers to take immediate action.

“After several years of working with Yogosha on Bug Bounty, we wanted to explore a different approach for these new applications. Access to the source code allowed us to take the analysis much further while working directly with our development teams throughout the engagement.”

Fabien Martinez, Cybersecurity Expert at Adevinta France

20 Repositories, 9 Specialized AI Agents and Three Layers of Validation

Augmented Pentesting is not about asking an AI to find vulnerabilities and treating its output as confirmed findings.

During this engagement, 9 specialized AI agents initially analyzed the 20 repositories. Each agent focused on a specific vulnerability category, rapidly expanding the scope of code coverage.

But this initial analysis was only the first step.

Every potential finding identified by the agents was then reviewed, challenged and enriched by Yogosha’s security experts. Finally, the experts manually examined the source code to confirm each vulnerability before documenting it.

AI agents expand coverage. Humans make the decisions.

This combination leverages AI’s ability to rapidly analyze large volumes of code without compromising what makes a pentest valuable: expert reasoning, contextual understanding and human validation.

“The goal of Augmented Pentesting is not to replace pentesters with AI agents. In this engagement, the agents allowed us to significantly expand coverage across 20 repositories, while our experts challenged and confirmed every finding. By enabling developers to implement fixes during the engagement, we also significantly reduced the time between detection, remediation and validation.”

Mohamed Arfaoui, Cybersecurity Expert at Yogosha

Turning Findings into Actionable Insights for Developers

Access to source code also changes how vulnerabilities are documented.

For every confirmed finding, Yogosha’s experts could pinpoint the affected file, line and code, provide actionable evidence and recommend a fix.

The objective was straightforward: to deliver reports that developers could act on directly, without requiring the security team to interpret them first.

In total, 62 vulnerability reports were submitted during the engagement.

But perhaps the most interesting metric lies elsewhere.

Find, Fix, Retest… All Within the Same Pentest

The engagement took place over a single weekend, with security and development teams working in parallel.

As soon as a vulnerability was identified and reported through the Yogosha platform, leboncoin’s developers could modify the source code and push a fix.

Yogosha’s experts would then retrieve the updated version, review the commit and check whether the vulnerability was still present in the code.

The process became a continuous loop:

Test → Finding → Fix → Retest → Validation

And this cycle could be repeated throughout the engagement.

By the end of the pentest, 18 vulnerabilities had already been addressed. Eleven had been closed on the platform, while seven others had been withdrawn directly by Yogosha after verifying the fixes. Of these 18 vulnerabilities, 16 had been fully remediated.

In other words, the pentest was no longer simply providing a snapshot of security at a given point in time: the security posture was already improving while testing was still underway.

“What made this engagement particularly interesting was that we didn’t have to wait until the end of the pentest to start fixing vulnerabilities. Findings were shared as they were identified, our developers could push fixes, and Yogosha would verify them directly in the source code. We were genuinely working in a continuous loop of detection, remediation and validation.”

Fabien Martinez, Cybersecurity Expert at Adevinta France

White-Box Pentesting: Going Faster and Deeper

White-box testing offers another significant advantage: starting with a detailed understanding of how the application actually works.

Rather than focusing solely on what can be exploited externally, security experts can examine authentication and authorization mechanisms, secrets stored in repositories, configurations, injection risks and potential data exposure.

Of the 55 findings that remained active at the end of the engagement, 26 related to authentication and authorization, while 14 involved secrets stored in repositories.

However, reviewing the source code also highlighted the limitations of white-box testing.

Some hypotheses can only be validated by interacting with a deployed application. Twelve areas for further investigation were identified for a potential second phase of testing against live applications, creating a natural bridge between white-box analysis and black-box validation.

From Point-in-Time Pentesting to a Continuous Security Loop

This engagement with leboncoin illustrates a broader evolution in offensive security.

The challenge is no longer simply about finding more vulnerabilities or finding them faster.

It is also about reducing the time between vulnerability discovery, understanding the issue, implementing a fix and validating the remediation.

AI provides the ability to review and analyze more code. Security experts bring offensive reasoning, context and validation. Developers can take immediate action on confirmed vulnerabilities.

And the platform brings all these stakeholders together within a single workflow.

Detect. Understand. Fix. Retest.

It is this combination of automation, human expertise and integration into existing team workflows that is gradually transforming point-in-time pentesting into a more continuous approach to security.