Company Overview
Founded in 2018 and headquartered in the United States, CalypsoAI is a startup focusing on artificial intelligence security. CalypsoAI’s market positioning is to open up a reliable path for the safe implementation of AI, solve the security threats faced by AI in enterprise applications, and ensure the safety and trustworthiness of AI models in reasoning. From the official website information [1], it can be seen that CalypsoAI complies with the compliance requirements of international authoritative certification and standard setting parties such as OWASP, NIST, and MITRE.
CalypsoAI provides AI protection capabilities to government, military, and enterprises, including the U.S. Department of Homeland Security, the U.S. Air Force, the Ministry of Aviation, etc. [2]. Forrester Research has listed CalypsoAI as an important AI security solution provider [3].
The core figures of the founding team of CalypsoAI include Neil Serebryany and Victor Ardulov.
As the co-founder of CalypsoAI, Neil Serebryany has a dual background in national security and venture capital. He has worked at the forefront of national security innovation in the Ministry of National Defense for a long time and is responsible for formulating AI application strategies in the defense field. He is widely regarded as a leader in this field. After founding CalypsoAI in 2018, Neil Serebryany led the company’s strategic direction and commercialization implementation, promoted the transformation of products from single-point tools to full-lifecycle AI security platforms, put forward the concept that “AI security needs to be deeply integrated with business needs”, and cooperated with Deloitte Middle East and other institutions to develop global enterprise-level AI security standards. Neil Serebryany’s career trajectory combines policy making and technology commercialization. He has invested in many AI startups as one of the youngest venture capitalists in the world and is well versed in the path of technology industrialization.
Victor Ardulov is one of the co-founders of CalypsoAI and worked at it from September 2018 to October 2022. Public information shows [4,5] that Victor Ardulov received his bachelor’s degree from the University of California, Santa Cruz, and then studied for his master’s and doctoral degrees at the University of Southern California. During his time at the University of Southern California, Victor Ardulov met Neil Serebryany and later served as Chief Scientific Officer for four years as a founding technical member of the team. He was responsible for developing machine learning model evaluation standards, software and tools, leading the development of underlying technologies such as adversarial attack detection and model robustness verification, and promoting the implementation of “zero trust AI architecture”. Victor Ardulov’s work focuses on introducing and standardizing best practices in machine learning and exploring various ways to incorporate measured model testing into products.
Figure 1 Neil Serebryany (left) and Victor Ardulov (right)
Product Introduction
To help network and AI teams discover and respond to potential risks in artificial intelligence, CalypsoAI has created an enterprise-level security protection layer-The Inference Perimeter, which provides unified security protection for large model reasoning across models and vendors regardless of the type of model used by the enterprise (Claude, GPT-4, etc.), model environment (public cloud, Private cloud, on-premises deployment), and corresponding suppliers (OpenAI, Meta, etc.).
As a platform for unified risk management of different AI models, The Inference Perimeter integrates three core modules: Red-Team, Defend and Observe to form an AI security protection system with proxy red team, real-time defense and automatic security execution functions.
Figure 2 The Inference Perimeter
1. Red-Team
CalypsoAI’s Red-Team module implements a red team strategy for GenAI, aiming to simulate real attack scenarios through systematic adversarial testing and stress testing, and discover and expose GenAI system vulnerabilities before being exploited by attackers.
Compared to the current red team strategies of other teams, CalypsoAI lists the following features:
1) Agentic Warfare: Simulate real adversarial interactions, enable models to participate in dynamic and adaptive dialogues, and explore deep-seated vulnerabilities and hidden risks that will appear in continuous interactions.
2) Extensive Signature Attacks: Use more than 10,000 continuously updated attack prompt libraries to systematically test the risks in model responses (such as prompt injection, incremental attacks, etc.).
3) Operational Attacks: Assess and detect vulnerabilities in the entire AI system from the infrastructure level, identify service paralysis risks, response delay vulnerabilities, and computing resource waste vulnerabilities.
4) Continuous Assessment: Establish automated adversarial testing to support repetitive continuous testing to maintain continuous governance of evolving AI models.
Figure 3 Red-Team Use Cases
2. Defend
Inference Defend intercepts threats before they invade AI applications, preventing rapid injection, jailbreaking and data leakage. With adaptive security and content moderation features, this module ensures AI safety compliance without affecting model performance.
The inference defend module has the following characteristics:
1) Inference-Layer Focus: Provide direct protection when AI models interact, and analyze each input and output of the model in real time to detect threats.
2) Customizable Controls: Allow enterprises to create customized security policies based on specific use cases, regulatory requirements and innovation strategies.
3) Immediate Protection: Pre-installed intelligent security scanning engine, automatically activated during deployment, to achieve zero-time difference parallel operation and rapid deployment of AI service launch and security protection.
4) Protect Not Obstruct: Innovate the “security resilience adjustment” mechanism, flexibly adjust security settings, and find the best balance between strictly adhering to the bottom line of security and stimulating business innovation.
Figure 5 Effect Demonstration of Defend Module
3. Observe
Inference Observe provides security teams with a panoramic, real-time AI security situation awareness interface to continuously track and monitor the use of AI models, detect security vulnerabilities, identify legal risks caused by non-compliant outputs, and ensure that AI systems operate as expected.
The characteristics of the inference observe service are:
1) Unified Monitoring & Reporting: Real-time AI security insights with detailed logging and audit trails provide security teams with the visibility they need to track, investigate and mitigate risks.
2) Global Dashboards: Centrally view AI usage and security events to ensure enterprise-level compliance.
3) Security & Risk Policies: An intelligent policy verification engine that automates the marking of AI risk applications that do not comply with corporate governance norms, regulatory requirements and endogenous security frameworks.
4) Seamless Integration: Smoothly connect SIEM, SOAR and work order systems to provide enterprise-wide visibility so that security teams can take action in their existing workflows.
Figure 5 Effect Demonstration of Observe Module
Overall, CalypsoAI’s The Inference Perimeter integrates three different protection functions: red team testing, observation perception and blocking defense into a unified platform. Compared with CalypsoAI’s previous two products Moderator and Vesper Validate for generative AI prompt content security scanning and model credibility verification, The current product line adjustment may be to cope with the gradually escalating full-link risk management and control needs, and to achieve a more comprehensive solution through technology integration, thereby building a security protection closed loop of “sensing-response-handling”. CalypsoAI’s current security scanning and non-compliance behavior marking have significant high automation features, which seamlessly connect with the existing workflow of customer enterprises to ensure the ease of use of the product. Its capabilities have also been confirmed in many government-enterprise cooperation. With its multiple advantages, it was shortlisted for the top ten RSAC innovation sandboxes in 2025 and was rated as a “leader in the AI security track.”
However, from the current product capabilities, CalypsoAI currently mainly covers general security frameworks. It may need to add additional relevant content to adapt to the compliance needs and security protection capabilities of some special industries (such as medical care and finance). Therefore, there is still room for improvement in compliance and security in vertical fields, and the maturity of the ecosystem needs to be further improved. In addition, although CalypsoAI’s official website emphasizes that its security protection covers the entire life cycle of AI, the current technical focus and application scenarios of core products are mainly focused on the reasoning layer. There are some blind spots in the coverage of AI system security threats, such as the lack of attack threat protection in the data preparation and processing stages and model training stages. Therefore, it is temporarily difficult to meet the full range of end-to-end security needs of enterprises from AI model development to application. The coverage of safety protection still needs to be further extended.
Summary
As a pioneer in the field of AI security, CalypsoAI has built real-time AI security situational awareness, achieved full-link risk coverage and dynamic security policy adjustment for AI applications, has the advantage of flexible deployment, and performs well in automated compliance inspections.
Faced with increasingly stringent regulatory trends, global companies are gradually incorporating the security protection capabilities and compliance construction of AI systems into their corporate strategies. From the iterative upgrade of the anti-attack defense system, to the construction of a full life cycle security management platform, to the systematic inspection and dynamic strengthening of compliance requirements, there is still broad room for development in the field of AI security protection.
References
[1] https://calypsoai.com/inference-platform/
[2] https://news.sohu.com/a/692352784_121649381
[3] https://www.chinaz.com/2023/0628/1537935.shtml
[4] https://contactout.com/Victor-Ardulov-83494640
[5] https://vardulov.github.io/
