Synthèse
To develop our SharePoint exploit chain, Rapid7 Labs undertook a research project divided into two main sprints, the first in January and the second in March, 2026. While both sprints did encompass more traditional vulnerability research such as manual code review and reverse engineering, a significant amount of the work was undertaken through an agent. Over 24 active days of agentic work, we leveraged 96 sessions, issued 256 prompts, and generated approximately 80,000 agentic tool calls. The in
To develop our SharePoint exploit chain, Rapid7 Labs undertook a research project divided into two main sprints, the first in January and the second in March, 2026. While both sprints did encompass more traditional vulnerability research such as manual code review and reverse engineering, a significant amount of the work was undertaken through an agent. Over 24 active days of agentic work, we leveraged 96 sessions, issued 256 prompts, and generated approximately 80,000 agentic tool calls. The initial January sprint was unsuccessful, resulting in no findings that could be leveraged for an exploit chain. We used this sprint to experiment with several different publicly available models, along with different workflows to navigate and reason across a massive and complex codebase.
However, our second sprint in March was successful and yielded, through a heavily prompted agent, a two-vulnerability exploit chain that achieved unauthenticated RCE. The improvement in quality between January and March in terms of agentic work, along with our improved workflows, was noticeable. This highlights the speed at which this field is evolving, how publicly available models are improving, and how as research teams develop their workflows, the results begin to compound.