
AI Model Security Testing and Adversarial Robustness
COMPANY OVERVIEW
TrojAI secures Artificial Intelligence systems through a comprehensive defense platform designed to protect models across their entire lifecycle, from development to deployment. Founded in New Brunswick, Canada in 2019, TrojAI focuses on governance and risk management for AI, providing two core products dedicated to securing enterprise-scale generative AI systems. TrojAI raised $5.75 million in seed funding in April 2024, led by Flying Fish with participation from Build Ventures, Techstars, Alteryx Ventures, and Flybridge Capital Partners.
CORE FOCUS
TrojAI enables safe enterprise adoption of machine learning and generative AI by addressing threat vectors unique to modern AI systems, including model poisoning, adversarial attacks, and behavioral exploitation. The dual-phase security architecture combines proactive vulnerability discovery with reactive threat mitigation — automated red teaming during development and AI-specific firewalling at runtime.
PRODUCTS & TOOLS
Detect — Autonomous Red Teaming — Proactive vulnerability discovery for AI models before deployment.
- Register models and configure attack libraries for comprehensive testing
- Autonomous red teaming agent that stress-tests models against known attack vectors
- Behavioral risk assessments to identify vulnerabilities before production
- Actionable remediation guidance for weaknesses revealed during testing
Defend — Runtime Security — Real-time monitoring and blocking for production AI systems.
- Design and enforce runtime security policies for deployed models
- Prompt injection detection and blocking in real-time
- Firewall deployment architecture for enterprise-scale protection
- PII detection and exfiltration prevention
MCP Protection — Securing agentic AI workflows and tool-use patterns.
- MCP Server protection against prompt injection attacks
- Register and monitor MCP servers across the enterprise
- Detect and block malicious tool-use patterns in agentic workflows













