As enterprises prepare to deploy hundreds of thousands of AI agents by 2028, manual security testing is becoming impractical. Vijil's new DART system aims to fill that gap by automating red-teaming with adaptive, multi-turn attacks that probe an agent's defenses in its own environment. Unlike static prompt-based tools, DART evaluates the target's responses, adjusts its tactics, and retries as many times as needed, covering the agent's tool-use, memory, and conversational behavior.
The tool ships with risk taxonomies based on OWASP, MITRE, and Vijil research, but can be customized to any enterprise risk catalog. It runs automatically under sustained load, with rate limiting and retries, and can be deployed via APIs as part of a CI/CD pipeline. For regulated environments, DART operates inside the customer's VPC or fully on-premises, producing auditable reports for both engineering and compliance teams.
In a benchmark evaluation using DecodingTrust-Agent, DART achieved an attack success rate 1.5 times that of its closest competitor, surpassing it in nine of twelve enterprise agent tasks spanning CRM, code, customer service, medical, research, and travel. The vendor notes that existing red-teaming services often run only days before deployment and rely on external consultants, making it hard for developers to fix issues in time. DART is part of the broader Vijil platform, which also includes modules for discovering shadow AI, enforcing production policies, and continuously improving agents after launch.