Vijil DART Automates Security Testing for Enterprise AI Agents
Vijil launches Diamond Adaptive Red Teaming for Agents (DART), an AI-powered tool that uses adversarial agents to find security flaws and policy violations in enterprise AI systems.
Vijil has introduced Diamond Adaptive Red Teaming for Agents (DART), a novel automated testing system engineered to proactively identify security vulnerabilities and policy violations within enterprise AI agents. DART distinguishes itself by employing its own suite of adversarial AI agents. These agents engage in multi-turn attacks, dynamically adapting their strategies across interactions to probe and test the target AI's defenses more effectively than traditional methods.
The proliferation of AI agents in enterprise environments is accelerating rapidly. Gartner predicts that by 2028, a typical global Fortune 500 company will deploy over 150,000 AI agents, a significant leap from fewer than 15 in 2025. This exponential growth presents a substantial challenge for AI engineering and governance teams, who often lack the necessary bandwidth and budget for comprehensive manual testing. Concurrently, threat actors are increasingly leveraging their own agents to launch sophisticated, sustained multi-turn attacks against these burgeoning AI systems.
Existing red-teaming tools often struggle to keep pace with the evolving threat landscape. Many rely on static test prompts that match known attack patterns but fail to develop novel strategies or evade target agent defenses. These tools typically focus on testing the underlying language model and chat interface, often operating outside the agent development lifecycle. This late-stage testing, frequently conducted days before deployment, leaves developers with insufficient time to address critical findings.
DART offers a fundamentally different approach. It empowers AI teams to test the entire agent, including its tool-use capabilities, memory functions, and multi-turn conversational behavior. The system utilizes adaptive attacks that mimic the target agent's operational environment, moving beyond fixed scripts. DART achieves a balance of speed, scale, and stealth, generating numerous multi-turn attacks designed to uncover and exploit the target agent's weaknesses. The system continuously evaluates the agent's responses, refines its tactics, and iterates as specified by the user, autonomously identifying vulnerabilities without explicit guidance.
Key features of DART include customizable risk coverage, drawing from established taxonomies like OWASP and MITRE, as well as enterprise-specific risk catalogs. Its adaptive attack chaining enables sophisticated probing of the attack surface across multiple turns and episodes. The tooling is developer-friendly, integrating with popular coding agent plugins and any agent framework or deployment platform, producing auditable reports for engineering and compliance teams. DART is also DevOps-ready, capable of running at scale within CI/CD pipelines and offering deployment flexibility, including on-premises and air-gapped environments.
In recent evaluations against the DecodingTrust-Agent benchmark, DART demonstrated a 1.5x higher attack success rate compared to its closest competitor, outperforming in nine out of twelve enterprise agent tasks across diverse domains such as CRM, code generation, customer service, medical, research, and travel.
"Your custom AI agent that can access your confidential data and take consequential action on its own requires a comprehensive and customized approach to quality control," stated Vin Sharma, CEO of Vijil. "There’s a big gap between an agent that appears ready in a demo and one that proves its reliability, security, and safety under pressure. DART closes that gap, saving weeks of effort and tens of thousands of dollars compared to manual red-teaming engagements, generic benchmarks, and open source prototypes."
DART is an integral part of the Vijil platform for evolving resilience into AI agents. Following DART's identification of agentic weaknesses, other Vijil modules perform root-cause analysis, implement policy-driven guardrails, and suggest code modifications. The broader Vijil platform encompasses modules for discovering shadow AI (Vijil Discover), identifying flaws before deployment (Vijil Diamond), ensuring production compliance (Vijil Dome), and continuous agent improvement (Vijil Darwin), collectively aiming to maintain agent resilience under adversarial conditions.