At Black Hat USA 2026, nearly 100 cybersecurity practitioners participated in a 48-hour event called SWARM, hosted by Tenable and sponsored by AWS, with technical staff from Anthropic serving as judges. The event focused on leveraging agentic AI to develop open-source defensive cybersecurity tools, which are now available on the CyberAgents Exchange. The initiative aimed to address the growing asymmetry in cyber capabilities, where agentic AI has significantly lowered the cost for attackers to find and exploit vulnerabilities, a trend not seen since the 1990s.
The core premise of SWARM was to empower defenders to build automation without requiring extensive development expertise, thereby closing a distribution gap that often sees security teams independently solving the same problems. Participants were tasked with identifying a key operational pain point and building a practical solution, which then had to be published open source to the CyberAgents Exchange to be eligible for prizes.
The projects developed at SWARM largely focused on "unglamorous" but critical tasks that consume significant time for security teams. These included agents for prioritizing findings, reconciling disparate asset inventories, correlating scanner outputs, and automating the verification of mitigations. Instead of building autonomous red teams or self-healing networks, the participants concentrated on creating connective tissue—tools that normalize data across existing systems, wrap current security tools for agent interaction, and package recurring analyst tasks into reusable skills.
One team created an agent that reads vendor advisories and mitigation notes, compares them against live findings, and recommends risk recasts, a task typically performed manually by senior engineers. Another developed an agent to correlate two different scanners to determine if a code flaw was actually reachable in a running application. A third focused on providing auditable evidence that a finding had already been mitigated.
The top three teams were recognized for their impactful contributions. First place went to Team Vauban for their "Chokepoint Finder," an MIT-licensed tool designed to address the "selection problem" faced by security teams overwhelmed with thousands of open findings but limited capacity for remediation. Chokepoint Finder, published as four agentic components (an agent, MCP server, skill, and playbook), ranks fixes rather than individual findings. It groups findings by the single action that resolves them (e.g., patches, base images, IAM roles) and then identifies the shortest ordered list of actions that mitigates the most weighted risk.
The Chokepoint Finder leverages its playbook to agentically inspect and rank fixes through eight distinct stages, incorporating a human decision point. It operates with a strict methodology: the agent holds no write credentials, evidence it cannot read or date results in a hold, and a re-scan is required to prove risk reduction before a record is closed. In a demonstration, the tool collapsed an estate with 3,734 synthetic findings across 783 assets into just seven concrete remediation actions. Each member of Team Vauban received $2,000 in Anthropic credits and an AWS specialist certification.
The winning ranking engine developed by Team Vauban is notable for its simplicity, shipping as a skill that runs solely on the Python standard library, requiring no additional packages, installation steps, or build pipelines. This demonstrates that practitioners can produce deployable solutions within a short timeframe, making such tools accessible for immediate use by colleagues. The overall objective of SWARM and the CyberAgents Exchange is to foster a collaborative environment where defensive cyber capabilities can compound, much like offensive tooling, by providing an easy way for defenders to share and leverage each other's work.
