The new DART platform automates security testing for autonomous agents by deploying adversarial models that learn and adapt to target defenses in real-time.
Vijil introduced Diamond Adaptive Red Teaming for Agents, or DART, on October 7, 2026, to address the expanding attack surface of enterprise-grade autonomous systems. The platform utilizes a fleet of adversarial agents designed to execute multi-turn, adaptive attacks against target AI, moving beyond the limitations of static prompt-based testing.
The rapid proliferation of agentic systems necessitates a shift in security validation methodologies. Gartner projections indicate that global Fortune 500 organizations will manage upwards of 150,000 agents by 2028, a significant increase from the fewer than 15 agents recorded in 2025. Manual red-teaming efforts cannot scale to meet this volume, leaving critical vulnerabilities exposed during the development lifecycle.
Traditional security tools often rely on static scripts that evaluate only the language model and chat interface. These legacy approaches fail to account for the complex, multi-turn behaviors inherent in modern agentic architectures, including tool-use and long-term memory. DART addresses these gaps by operating within the agent’s environment to probe for weaknesses across various operational contexts.
The system integrates directly into development workflows via plugins for coding environments like Claude Code and Codex. It supports automated CI/CD pipelines, allowing security teams to implement rate-limiting, retries, and role-based model configurations. This automation ensures that security testing remains consistent as agents evolve through continuous deployment cycles.
Deployment flexibility remains a core design tenet of the platform. Organizations can host DART within their own virtual private clouds or in air-gapped, on-premises environments to satisfy stringent regulatory requirements. This capability allows sensitive data to remain isolated while the system performs its adversarial probes.
DART utilizes iterative feedback loops where the adversarial agents adjust their tactics based on the target’s responses. This autonomous exploration allows the system to discover vulnerabilities without requiring predefined attack vectors. By simulating the strategies of sophisticated threat actors, the platform provides a realistic assessment of an agent’s resilience under sustained pressure.
Performance metrics underscore the efficacy of the adaptive approach. In evaluations utilizing the DecodingTrust-Agent benchmark, DART achieved an attack success rate 1.5 times higher than its closest competitor. The system outperformed alternatives in nine of twelve enterprise tasks, including domains such as CRM, medical research, and automated travel planning, proving its superior capability in identifying complex logic flaws.
Vin Sharma, CEO of Vijil, emphasized that the transition from a functional demo to a production-ready agent requires rigorous quality control. He noted that DART reduces the time and financial investment typically associated with manual red-teaming engagements by identifying flaws before deployment. The platform functions as a component of the broader Vijil Diamond ecosystem, which facilitates root-cause analysis and the implementation of policy-driven guardrails.
The integration of DART into the development lifecycle provides engineering teams with granular, auditable reports that bridge the gap between security findings and compliance requirements. By generating actionable data on agent behavior, the system enables developers to refine guardrails and modify code before the agent interacts with sensitive enterprise data. This proactive approach mitigates the risk of unauthorized tool execution or memory leakage in production environments.
Future development of the Vijil ecosystem aims to bridge the gap between initial vulnerability discovery and active remediation. Once DART identifies a flaw, other modules within the platform suggest specific code modifications to strengthen the agent’s defenses. These integrated workflows are intended to ensure that security measures keep pace with the rapid iteration cycles characteristic of modern AI development.
The adoption of automated red-teaming systems forces a recalibration of security budgets within enterprise AI governance frameworks. Organizations that integrate these adaptive testing modules into their CI/CD pipelines demonstrate a measurable reduction in post-deployment vulnerabilities compared to those relying on static, manual evaluation methods.
