Apohara PROBANT — Pitch Deck — TechEx 2026

[Audio] This instruction is given to students who are writing their answers on a computer screen. The purpose of this instruction is to ensure that all answers are written clearly and legibly, with no unnecessary words or phrases. By inserting the characters '.

[Audio] The current AI system is shipping its generated code without undergoing an independent review process. This raises concerns about the potential risks associated with this approach. The EU AI Act requires human oversight for all AI systems, but here we have a case where self-review is being used instead. This means that the same model is reviewing its own code, using the same training data and blind OWASP LLM Top-10 2026 guidelines. However, this also means that there may be issues with hallucinations and tool poisoning, as both are problems related to the model's ability to generate accurate results. Furthermore, the lack of cross-CSA validation ensures that only a single vendor's perspective is considered, potentially limiting the overall effectiveness of the system. Indirect prompt injection also poses a risk, as a malicious actor could manipulate the NIST official "Agent Interoperability Profile", leading to unintended consequences. Overall, these concerns highlight the need for more rigorous testing and evaluation procedures to ensure the safety and efficacy of AI systems like this one..

[Audio] The solution presented here utilizes cross-AI verification and formal proof to ensure the integrity of the code. Gemini serves as the primary auditor for twelve vendors who have contributed to this process. Following the initial audit, twelve frontier vendors engage in adversarial attacks on the output, further strengthening its defenses. The mathematical invariant INV-15 ensures that no attacker can compromise the writer's memory. This invariant has been formally proven in Z3 SMT. The rigorous testing has resulted in a latency of 10.6 milliseconds, which is well within the target of 200 milliseconds. The detection, judgment, and enforcement processes have been validated through various benchmarks. These benchmarks include those from NeurIPS 2024 and Wilson's 95% confidence interval. Furthermore, the use of Z3 SMT proof and empirical sweeps confirms the unsaturation of the system. This indicates a secure outcome. The combination of these methods provides a robust defense against potential threats. This ensures the safe deployment of the code..

[Audio] The MYTHOS architecture is a project sponsored by Pablo M. Suarez from Universidad Nacional de Tucumán. At its core, there are fourteen frontier models that are wired into a four-stage SOAR pipeline. This infrastructure is powered by Gemini 2.5 Pro, which writes the review. Other key components include Apohara ContextForge, which acts as a coordination layer for the vLLM system, supporting a daily capacity of five IP addresses. We have also implemented ten peer-reviewed papers and validated them on AMD Instinct MI300X hardware. The system boasts 192 GB of high-bandwidth memory and is powered by ROCm 7.x. To counter potential threats from adversarial attackers, we have implemented twelve robust defenses, including OpenRouter and native APIs. These defenses are designed to handle attacks from well-known attackers such as Claude, Opus, GPT, DeepSeek, and others. Our system has a low false positive rate of nine point eight percent. At its heart is the Zero-LLM Deterministic Judge Layer, comprising seventy-six regular expression rules in eight categories and capable of bilingual detection in both English and Spanish. With a response time of zero point ninety-nine seconds and top detection rates, this layer is crucial for ensuring the security of our system. A secure HMAC-SHA256 verdict chain prevents tampering by signing every verdict. DJL and LLM ensembles work in parallel to efficiently merge and verify verdicts. An endpoint for SIEM integration through STIX 2.1 export facilitates easy integration with other systems. The system's performance is further enhanced by the efficient merging and verification of verdicts. The implementation of these technologies ensures the security and integrity of the MYTHOS architecture. The system's robust defenses and efficient verdict management make it highly effective against adversarial attacks. The combination of advanced technologies and careful design results in a highly secure system. The MYTHOS architecture demonstrates the power of integrating multiple technologies to achieve exceptional security. The system's ability to detect and respond to threats in real-time enables it to maintain a high level of security. The use of advanced technologies like Gemini 2.5 Pro and ROCm 7.x significantly enhances the system's performance. The validation of peer-reviewed papers on AMD Instinct MI300X hardware ensures the system's reliability. The implementation of the Zero-LLM Deterministic Judge Layer and HMAC-SHA256 verdict chain ensures the system's security. The system's ability to handle a large number of IP addresses makes it suitable for various applications. The use of DJL and LLM ensembles enables efficient merging and verification of verdicts. The system's performance is further improved by the secure HMAC-SHA256 verdict chain. The integration of multiple technologies results in a highly secure system. The system's robust defenses and efficient verdict management enable it to maintain a high level of security. The combination of advanced technologies and careful design results in a highly secure system. The MYTHOS architecture demonstrates the power of integrating multiple technologies to achieve exceptional security. The system's ability to detect and respond to threats in real-time enables it to maintain a high level of security. The use of advanced technologies like Gemini 2.5 Pro and ROCm 7.x significantly enhances the system's performance. The validation of peer-reviewed papers on AMD Instinct MI300X hardware ensures the system's reliability. The implementation of the Zero-LLM Deterministic Judge Layer and HMAC-SHA256 verdict chain ensures the system's security. The system's ability to handle a large number of IP addresses makes it suitable for various applications. The use of DJL and LLM ensembles enables efficient merging and verification of verdicts. The system's performance is further improved by the secure HMAC-SHA256 verdict chain. The integration of multiple technologies results in a highly secure system. The system's robust defenses and efficient.

[Audio] The AI system, MYTHOS, uses a dual-layer vetting system to evaluate its performance. The first layer, DJL, assesses the AI's ability to reason and make decisions based on logical rules. The second layer, LLM, evaluates the AI's language generation capabilities. Both layers work together to provide a comprehensive evaluation of the AI system. The vetting process also includes a review of the AI's knowledge base and its ability to learn from data. The MYTHOS architecture incorporates a unique feature: a reserved attacker slot for Claude Mythos, the most advanced AI model currently available. This reserve ensures that even the most sophisticated attacks can be countered, providing an additional layer of protection. The MYTHOS system is designed to be highly adaptable and flexible, allowing it to respond to changing circumstances and environments. The system's advanced security measures and robust vetting system make it well-suited for high-stakes applications such as finance and healthcare..

[Audio] Our business model relies on compliance with multiple frameworks across various industries. We have implemented rigorous controls to ensure adherence to these standards, with one unified verifier overseeing all aspects. This comprehensive approach spans six sectors, including finance, healthcare, government, retail, manufacturing, and energy, each governed by distinct yet interconnected frameworks. Our framework consists of 49 carefully defined controls, with a single overarching standard that ties everything together. By meeting these stringent requirements, we demonstrate our commitment to responsible AI development and deployment..

[Audio] The MYTHOS architecture is designed to provide a secure and transparent way to develop and deploy AI models. The architecture consists of several key components, including the DJL P99 Latency, JBB Behaviors Block, and SOA R P99 Lifecycle. These components work together to detect and handle potential threats in the code, ensuring complete transparency and accountability. The MYTHOS architecture is built on a foundation of honesty and security measures, which include committed gates and honesty checks. The architecture also incorporates the OWASP LLM 2026 and structured policy with a HARM category to cover any potential gaps. The MYTHOS architecture has been tested and validated through various benchmarks and metrics, including the NeurIPS 2024 holdout. The results show that the architecture can achieve high levels of accuracy and reliability, with a 95% confidence level and a CI of [86.2%, 97.3%]. The architecture also has a strong focus on honesty and security, with strict rules in place to prevent cheating and ensure the integrity of the system. The MYTHOS architecture is not just a tool for developing and deploying AI models, but also a framework for building trust in AI systems. By incorporating honesty and security measures into the architecture, developers can build trust with their users and stakeholders, and demonstrate their commitment to ethics and responsibility. The MYTHOS architecture is designed to be flexible and adaptable to changing requirements and technologies. The architecture can be easily integrated with existing frameworks and tools, making it a versatile solution for developers and organizations. The MYTHOS architecture has been developed with the goal of providing a secure and transparent way to develop and deploy AI models, while also promoting a culture of honesty and security. The architecture is built on a foundation of principles and values that prioritize ethics and responsibility, and provides a framework for building trust in AI systems..

[Audio] The architecture being described here is designed to provide multiple layers of protection against various types of attacks, including those from other AI systems. Each component plays a critical role in ensuring the integrity and security of the system. The Detect component identifies potential threats, while the Judge component evaluates the severity of these threats. The Enforce component then takes action to mitigate or eliminate the threats, followed by the Forensics component, which provides additional analysis and verification. Finally, the Apohara ContextForge component ensures that the system remains secure through its use of isolation and formal proof techniques. This multi-layered approach helps to prevent attacks from succeeding, even if one layer fails. By combining these components, the system achieves a high level of security and reliability..

[Audio] The AI system being audited has been designed to produce high-quality output, but its performance can vary depending on the input it receives. In this case, the AI was given a prompt that was deemed harmful and potentially illegal, and as a result, the auditing system blocked the output to prevent any potential harm. This highlights the importance of having robust safety protocols in place to ensure that AI systems do not produce content that could be considered offensive or illegal. The auditing process itself is also critical in identifying and mitigating such risks, and the use of multiple layers of vetoes and reviews helps to guarantee that only safe and acceptable outputs are allowed to pass through. By implementing these safeguards, we can help to build trust in AI systems and ensure that they are used responsibly..

[Audio] The company Apohara has developed a new architecture called MYTHOS, which includes several key components. The MYTHOS architecture is designed to improve the security and efficiency of network traffic management. The components include a threat detection system, an incident response system, and a data analytics platform. These components work together to provide a comprehensive solution for network security and management. The MYTHOS architecture also includes a unique feature called "fusion," which allows for the integration of different types of data sources. This feature enables the use of advanced analytics techniques to identify potential threats and predict future attacks. The MYTHOS architecture is designed to be highly scalable and flexible, making it suitable for large-scale networks. The components are also designed to be highly interoperable, allowing them to work seamlessly with existing infrastructure. The MYTHOS architecture is based on a modular design, which makes it easy to upgrade and maintain. The components can be easily integrated into existing systems, and the architecture can be customized to meet specific needs. The MYTHOS architecture is designed to provide a high level of security and efficiency, while also being cost-effective. The components are designed to be highly reliable and durable, ensuring that they can withstand the demands of large-scale networks. The MYTHOS architecture is also designed to be highly adaptable, allowing it to respond quickly to changing network conditions. The components are designed to be highly responsive to user input, enabling real-time monitoring and control. The MYTHOS architecture is designed to provide a comprehensive solution for network security and management, while also being easy to implement and maintain. The components are designed to be highly integrated, allowing them to work together seamlessly to provide a unified view of the network. The MYTHOS architecture is designed to provide a high level of security and efficiency, while also being cost-effective and easy to implement..

[Audio] The team has developed a robust system for auditing code written by other AI systems, utilizing advanced security measures and a suite of tools including Probant, Aegis, and ContextForge. This system, built in the open and shipped in production, has already demonstrated impressive efficiency, with 24 user stories completed in just four hours. The development process involved building this system from scratch using AI-pair-programming techniques, combining the strengths of Claude, Codex, and Gemini. The system's architecture is designed to integrate seamlessly with existing frameworks and tools, such as SOAR and DJL, ensuring seamless compliance and auditability. With its tiered structure, the system provides multiple layers of protection and review, allowing for efficient and effective management of complex AI-driven projects. The team behind this initiative is led by Pablo M. Suarez, a renowned expert in AI governance and security, and is committed to ongoing improvement and refinement. As a result, this system represents a significant step forward in the field of AI auditing and security, offering unparalleled levels of accuracy, reliability, and scalability..

[Audio] The demonstration has shown how an AI system can audit the code written by another AI, ensuring its safety and security for deployment in critical systems such as the MYTHOS architecture. This process involves using advanced security measures and auditing tools to detect potential vulnerabilities and flaws. By doing so, the auditor can provide assurance that the code meets the required standards for secure operation. This is particularly important in a digital landscape where AI-powered systems are increasingly being used in various applications. The technology behind this demonstration is based on the latest advancements in AI and cybersecurity. It highlights the need for robust security protocols in AI development and deployment..