TrojAI
Securing AI Innovations

James Stewart, CTO

These advancements—whether predictive AI or the newer generative AI—push the boundaries of innovation, helping companies improve organizational efficiency and gain competitive advantage.

However, a critical concern is growing amid these developments: the security of AI innovations.

“These advancements, while offering immense potential, also introduce vulnerabilities that malicious actors can exploit,” says James Stewart, CTO of TrojAI.

These AI models and applications, which may lack robust security measures, can become entry points for cyberattacks, hindering business operations and leading to financial and reputational risks. Businesses need to thoroughly pentest and monitor their AI models and applications to mitigate these risks and safeguard their operations.

An AI security company, TrojAI helps large enterprises navigate the fast-changing threat landscape at scale. Its strategy involves combining a traditional cybersecurity mindset and specialized deep learning to address the unique challenges in AI security.

Tackling the Challenges

The first challenge for AI security lies within itself. Every system has distinct competence bounds, leading to different cybersecurity vulnerabilities.

Data scientists train AI models with a limited dataset and reserve a portion for evaluating model performance. Since the training and testing data naturally belong to the same distribution and share similar characteristics, the resulting model is susceptible to manipulation through various unpredictable scenarios and edge cases. These unpredictable edge cases account for a significant proportion of situations that present to the model when deployed in the real world.

TrojAI addresses this inherent limitation with specialized penetration tests and a centralized AI firewall to detect and defend against potential issues. It offers two comprehensive solutions— TROJAI::DETECT and TROJAI::DEFEND—to manage the complexities of AI and LLM solutions by both surfacing and protecting models’ deficiencies and vulnerabilities.

Understanding Model Vulnerabilities

The TROJAI::DETECT platform helps enterprises identify potential vulnerabilities and threats. It contains an extensible library of more than 100 out-of-the-box security tests and automates penetration testing across a range of AI disciplines.

This platform artificially generates natural and malicious input perturbations to assess model performance under various attack scenarios, such as data poisoning, model evasion and model extraction. TROJAI::DETECT is purpose built for cybersecurity practitioners, mapping results back to the leading risk frameworks. These results provide actionable insights into the weaknesses of each model.
The Active Defender

TROJAI::DEFEND acts as a centralized firewall for the enterprise. With security controls in one place, the platform offers both real-time and out-of-band protection by identifying, flagging and blocking inferences or requests that could cause issues using an AI-intelligent rules engine.

TROJAI::DEFEND monitors specific weaknesses uncovered during penetration testing while monitoring for emerging threats, including adversarial prompts, hallucinations and denial of service attacks. Its performance can scale as new use cases are introduced, making this centralized strategy far superior to other API approaches. In addition, an extensible rules engine allows TrojAI to continually create new detections to protect the models against evolving threats.

"We monitor emerging regulations and actively contribute to creating AI standards. Our out-of-the-box protections, aligned with threat landscapes such as OWASP Top 10's and MITRE Atlas, and trust frameworks like Gartner's AI trust, risk and security management (AI TRiSM), provide cybersecurity teams with the actionable tools they need," says Stewart.

TrojAI’s platform can accommodate clients with an extensive AI portfolio—whether traditional discriminative AI/ML or emerging generative AI models. It effectively manages high demand and monitors millions of tokens daily to safeguard AI systems. For example, the platform can detect, flag, obfuscate, or block personally identifiable information (PII), copyright issues, intellectual property leaks, toxicity, hallucinations, adversarial prompts, and denial of service attacks, which could compromise model behavior.

TrojAI aims to enable enterprises to innovate safely through a comprehensive cybersecurity platform that safeguards existing and ongoing AI investments.

To achieve this goal, the company focuses on scalability and configurability, which are at the core of its AI security solutions. Its platform is deployed on-premises to ensure all data stays within the client's environment. Through its fully containerized architecture, authentication and encryption capabilities, and ability to integrate with existing security infrastructure, event management (SIEMs), and data lakes, TrojAI provides an attractive, comprehensive security layer for enterprise innovations.

A recent success story involves a Fortune 100 client who needed the ability to assess and protect model risk across hundreds of heterogeneous models and GenAI applications.

Data scientists often train AI models with a limited dataset and reserve a portion for evaluating model performance. Since the training and testing data belong to the same distribution and share similar characteristics, the resulting model is susceptible to manipulation through various scenarios and edge cases

With TrojAI’s automated pentesting and post-deployment monitoring, the client gained clear visibility into the inherent and residual risks on a model-by-model basis, eliminating data science ‘black boxes.' This enabled their cybersecurity team to have meaningful discussions about AI security. The platform provided their team with actionable risk assessments and real time protections, which helped the client improve their AI security posture management—all toward supporting rapid innovation.

The TrojAI Engine

A team of tenured professionals with data science and cybersecurity expertise fuels TrojAI’s success. Under its new CEO, Lee Weiner, the company aims to augment its product capabilities. This industry-leading platform and real-world expertise make TrojAI a top choice for AI security.

AI and ML will continue to evolve and shape tomorrow's technology. With TrojAI as an AI security partner, enterprises can confidently adopt AI innovations.