The new Diamond Adaptive Red Teaming framework provides an automated, multi-turn approach to identifying security vulnerabilities in autonomous enterprise AI systems.
Vijil has introduced Diamond Adaptive Red Teaming for Agents, a specialized testing framework designed to evaluate the security posture and policy adherence of enterprise-grade AI agents. Released on October 7, 2026, the system replaces static prompt-based testing with a dynamic, multi-turn adversarial approach that mimics the behavior of sophisticated threat actors.
The architecture of DART employs a fleet of adversarial agents that probe the target system across multiple turns and engagement episodes. This methodology allows the system to learn and adapt its tactics in real-time, effectively exploring the target agent’s memory, tool-use capabilities, and decision-making logic. Unlike traditional red-teaming services that rely on predefined prompt sets, DART autonomously identifies vulnerabilities without requiring explicit instructions on which attack vectors to prioritize.
Integration with existing development pipelines is facilitated through plugins for coding environments such as Claude Code and Codex. The system is designed to operate within a customer’s own Virtual Private Cloud or air-gapped infrastructure, ensuring that sensitive data remains localized during the testing process. This deployment flexibility addresses the requirements of highly regulated industries where external cloud-based testing services may be prohibited.
Operational efficiency is achieved through a CI/CD-ready workflow that supports automated, high-load testing. The system provides auditable reports that bridge the gap between engineering teams and compliance officers, documenting both the identified flaws and the context of the adversarial maneuvers. This automated feedback loop is intended to replace manual red-teaming engagements, which often occur too late in the development lifecycle to permit substantive architectural remediation.
Performance benchmarks highlight the efficacy of this adaptive approach, with DART achieving an attack success rate 1.5 times higher than competing solutions in tests using the DecodingTrust-Agent benchmark. The system demonstrated superior performance across twelve distinct enterprise domains, including CRM, medical, and code generation tasks. These results suggest that static evaluation methods are increasingly insufficient for detecting the complex failure modes inherent in autonomous agentic systems.
The technical superiority of DART is rooted in its ability to chain attacks across disparate episodes, a capability that mimics the persistence of real-world adversaries. By maintaining state across multiple interactions, the system can bypass defenses that only monitor for single-turn malicious inputs. This depth in stealth allows the framework to uncover edge-case vulnerabilities that remain hidden during standard unit testing or basic interface validation.
Vin Sharma, CEO of Vijil, emphasizes that the transition from demonstration-ready prototypes to production-hardened agents requires a shift toward continuous quality control. The platform aims to mitigate the risks associated with agents that possess the authority to access confidential datasets and execute consequential actions independently. By identifying weaknesses early, the system provides a mechanism for developers to implement policy-driven guardrails before deployment.
The broader Vijil platform provides a comprehensive lifecycle management approach, with DART serving as the primary diagnostic module. Once vulnerabilities are surfaced, subsequent modules facilitate root-cause analysis and suggest specific code modifications to remediate the identified flaws. This integrated ecosystem is designed to maintain agent resilience against evolving attack methodologies as new models and user interactions are introduced into the production environment.
The rapid proliferation of enterprise agents, which Gartner projects will scale to over 150,000 per Fortune 500 company by 2028, necessitates a move away from manual oversight. Organizations are currently struggling to maintain security standards as the volume of deployed agents outpaces the capacity of human governance teams. DART targets this scalability bottleneck by providing a repeatable, automated testing capability that evolves alongside the threat landscape.
Future development will likely focus on expanding the taxonomy of supported risks and enhancing the depth of the adaptive attack chains. As threat actors refine their own agent-based offensive strategies, the ability to conduct sustained, multi-turn defensive testing will become a critical component of AI infrastructure. The industry will be monitoring how these automated red-teaming systems influence the standardization of safety protocols across complex agentic deployments.
