EPAM

Excellencies and Co-Chairs. I speak today from my vantage point as a LEADER at the intersection of Cloud AI and security. For those of us tasked with SCALING these technologies, AI is not a monolith; it is a complex stack of compute, data, and models that requires a 'SECURE-BY-DESIGN' architecture to be viable. We welcome this Global Dialogue as the essential platform to move beyond high-level ethics toward a HARMONIZED, TECHNICAL FRAMEWORK for international cooperation. Safety cannot be a 'bolt-on' feature; it must be the FOUNDATION upon which the entire ecosystem is built." "The primary value of AI lies in its ability to solve systemic global challenges, but this potential is GATED by trust. To unlock the 'good,' our scientific assessments must prioritize THREE technical pillars: First: From Theory to Resilience. We must transition from philosophical safety to the development of TRUSTWORTHY AI systems that are measurably resilient against adversarial attacks, such as data poisoning and model inversion. Second: Open-Source as a Leveler. Transparency is the ULTIMATE security audit. Promoting open-source models is critical—it allows developing nations to audit, adapt, and secure tools within their own contexts preventing 'security dependencies.' Third: The Compute Prerequisite. We cannot bridge the digital divide without addressing the COMPUTE GAP. Facilitating access to high-performance computing is not just an economic issue; it is a SECURITY NECESSITY. Without it, capacity-building remains a theoretical exercise." "Regarding the roadmap for our July discussions, the Scientific Panel must provide the 'TECHNICAL TELEMETRY' for three priority areas: Interoperability is PRIORITY ONE. Fragmented governance is a security risk. We need a 'common technical language' for AI risk so that security standards remain CONSISTENT across borders. Evidence-Based Benchmarks. Policy without technical telemetry is just GUESSSWORK. The Independent Panel must deliver benchmarks that distinguish between marketing hype and VERIFIABLE system capabilities. Red-Teaming and Accountability. We should cluster discussions around ROBUST HUMAN OVERSIGHT to ensure that transparency is a verifiable mechanism for compliance with international law." "The first Global Dialogue should aim to deliver TWO practical, 'Day One' outputs: A Global AI 'Red-Teaming' Network: A collaborative platform where we share vulnerability data and threat models for LLMs in REAL-TIME. The 'Interoperability Stack': A set of technical standards that allow national frameworks to 'TALK' to one another. This ensures that 'trustworthy' in one jurisdiction doesn't become a SECURITY LIABILITY in another. Let's ensure that 'safe and secure' isn't just a goal, but our VERIFIABLE STATE. Thank you."

• Safe, secure and trustworthy AI
• AI capacity-building
• Open-source software, open data and open AI models
• Protection and promotion of human rights

Our priorities center on the transition from high-level ethical theory to a harmonized, technical framework. Safe, Secure and Trustworthy AI: We view safety not as a "bolt-on" feature but as the fundamental "secure-by-design" architecture necessary for any viable AI ecosystem. Our focus is on making systems measurably resilient against specific technical threats like data poisoning and model inversion.AI Capacity-Building: Real-world capacity-building is a security necessity, not just an economic goal. Addressing the "compute gap" is the prerequisite for moving beyond theoretical exercises to functional, secure participation in the global AI economy. Interoperability of Governance: Fragmented governance creates inherent security risks. We advocate for a "common technical language" and an "Interoperability Stack" -- a set of standards that allow national frameworks to communicate, ensuring that "trustworthy" status remains consistent across borders rather than becoming a liability. Open-Source Models: We view transparency as the ultimate security audit. Promoting open-source models is critical for preventing "security dependencies," as it empowers developing nations to audit and secure tools within their own unique contexts. By focusing on these pillars, the Global Dialogue can provide the "technical telemetry" needed to distinguish between marketing hype and verifiable system capabilities.

A critical emerging issue is the need for Real-Time Technical Telemetry and Shared Threat Intelligence. Current governance often relies on static assessments, but the rapid evolution of LLMs requires a Global AI "Red-Teaming" Network. This would facilitate the real-time sharing of vulnerability data and threat models across international lines. Without this collaborative infrastructure, international law and national safety standards will struggle to keep pace with verifiable mechanisms for compliance. Furthermore, we must address the distinction between "philosophical safety" and "adversarial resilience". Governance must evolve to include evidence-based benchmarks that measure how a system performs under active attack, ensuring that "safe and secure" is a verifiable state rather than just a policy goal.

Governance gaps in these thematic areas significantly impact the Cloud AI and security sector by creating technical and operational friction. Currently, fragmented governance acts as a primary security risk; without a "common technical language" for AI risk, security standards remain inconsistent across borders, making it difficult to maintain a unified defense against global threats. Furthermore, the "compute gap" prevents the digital divide from closing, rendering capacity-building a "theoretical exercise" rather than a functional reality for many regions. In the absence of evidence-based benchmarks, the sector often struggles to distinguish "marketing hype" from "verifiable system capabilities," leading to policy based on guesswork rather than "technical telemetry". Recent developments are shifting the focus toward "secure-by-design" architectures where safety is a foundation, not a "bolt-on" feature. Advances in adversarial resilience are helping the sector move from philosophical safety to systems that are measurably protected against data poisoning and model inversion. Finally, the push for open-source models is gaining momentum as a tool to prevent "security dependencies," allowing nations to audit and secure tools within their own contexts. To bridge remaining gaps, the sector is moving toward real-time collaboration, such as Global Red-Teaming Networks and "Interoperability Stacks," ensuring that "trustworthy" status in one jurisdiction does not become a security liability in another.

The AI Dialogue serves as the essential platform to transition international cooperation from high-level ethical discourse toward a harmonized, technical framework. By moving beyond philosophical safety, it enables the global community to establish a "common technical language" for risk, ensuring that security standards remain consistent across borders. According to the framework, the Dialogue can advance cooperation through three specific roles: Establishing Technical Telemetry: The Dialogue's Scientific Panel can deliver evidence-based benchmarks that distinguish between marketing hype and verifiable system capabilities, providing a factual basis for international policy. Creating an Interoperability Stack: It can facilitate the development of technical standards that allow diverse national frameworks to "talk" to one another, preventing a "trustworthy" status in one region from becoming a security liability in another. Building Collaborative Networks: The Dialogue can launch practical, "Day One" deliverables like a Global AI "Red-Teaming" Network. This platform allows nations to share vulnerability data and threat models for Large Language Models (LLMs) in real-time, fostering a culture of collective resilience rather than isolated defense. Ultimately, the Dialogue ensures that capacity-building is treated as a security necessity rather than a theoretical exercise. By addressing the compute gap and promoting open-source models, the Dialogue empowers developing nations to audit and secure tools within their own contexts, preventing "security dependencies" and ensuring transparency acts as the ultimate security audit.

The AI Dialogue should build upon existing international efforts such as the G7 Hiroshima AI Process, the Bletchley Declaration, and the UN High-Level Advisory Body on AI. It should also connect with technical standard-setting bodies like ISO/IEC and IEEE to ensure governance is rooted in a "common technical language" rather than just policy. The added value of the AI Dialogue lies in its ability to move beyond high-level ethics toward a harmonized, technical framework for international cooperation. While other forums focus on principles, this Dialogue can deliver the "technical telemetry" required to distinguish "marketing hype" from "verifiable system capabilities". Specific practical contributions include: A Global AI "Red-Teaming" Network: A real-time collaborative platform for sharing vulnerability data and threat models for Large Language Models (LLMs).The "Interoperability Stack": A set of technical standards that allow national frameworks to "talk" to one another, preventing a security status in one jurisdiction from becoming a liability in another. Addressing the Compute Gap: Elevating compute access from an economic issue to a security necessity, ensuring capacity-building is more than a "theoretical exercise". Open-Source Promotion: Providing a mechanism for developing nations to audit and secure tools within their own contexts, which serves as the ultimate security audit. By providing this roadmap, the Dialogue ensures that "safe and secure" becomes a verifiable state for the entire global ecosystem.

Stakeholders can contribute by shifting the focus from philosophical safety to adversarial resilience and verifiable compliance. Contributions should prioritize the following roles: Technical Experts & Industry: Provide the "technical telemetry" and benchmarks needed to distinguish marketing hype from verifiable system capabilities. Developing Nations: Lead the adaptation of open-source models to ensure tools are secured within local contexts, preventing "security dependencies". Scientific Panels: Develop an "Interoperability Stack"—technical standards that allow disparate national frameworks to "talk" to one another. Recommendations for Format and Structure The Dialogue should move beyond high-level ethics toward a harmonized, technical framework built on these structural pillars: Modular Working Groups: Cluster discussions around Robust Human Oversight and Red-Teaming to ensure transparency is a verifiable mechanism for international law. A "Secure-by-Design" Track: Dedicate a stream to transitioning from theory to resilience, focusing on protecting against data poisoning and model inversion. Real-Time Collaborative Platforms: Establish a Global AI "Red-Teaming" Network as a "Day One" output to share vulnerability data and threat models in real-time. Resource-Equitable Infrastructure: Structure the roadmap to treat compute access as a security necessity, ensuring the "compute gap" is addressed to make capacity-building viable. This structure ensures that "safe and secure" is not merely a policy goal, but a verifiable state across the entire global ecosystem.

Based on the framework of securing the AI ecosystem, the following perspectives are currently underrepresented and require active integration: Underrepresented Perspectives Developing Nations and the "Compute Gap": These regions are often excluded from high-level governance because capacity-building remains a "theoretical exercise" without access to high-performance computing. Open-Source Contributors: Local developers in diverse contexts are essential for preventing "security dependencies," yet their role in providing the "ultimate security audit" is frequently sidelined. Technical Red-Teaming Experts: Beyond policy makers, the voices of those conducting "measurably resilient" testing against adversarial attacks—like data poisoning—are needed to move from philosophical safety to verifiable states. Mechanisms for Inclusion A Global AI "Red-Teaming" Network: Establishing a collaborative platform for sharing vulnerability data in REAL-TIME would allow technical experts from all regions to contribute to global safety standards. The "Interoperability Stack": Creating technical standards that allow national frameworks to "TALK" to one another ensures that "trustworthy" status in one jurisdiction does not become a security liability in another. Compute as a Security Necessity: Global discussions must pivot to treat access to compute as a prerequisite for secure innovation, ensuring developing nations can audit and secure tools within their own contexts. Evidence-Based Benchmarks: Moving toward "TECHNICAL TELEMETRY" ensures that governance is based on verifiable system capabilities rather than "marketing hype," allowing smaller players to be judged on technical merit. By focusing on these practical deliverables, the AI Dialogue can ensure that "safe and secure" is a universal, VERIFIABLE STATE.

To move beyond static panels, the AI Dialogue should adopt formats that prioritize "technical telemetry" and real-time collaboration. The following innovative engagement formats would foster a more dynamic and verifiable dialogue: Live "Red-Teaming" Simulations: Instead of theoretical debates, sessions should feature live demonstrations of model vulnerabilities, such as data poisoning or model inversion. This shifts the focus toward creating measurably resilient systems. "Interoperability Hackathons": Technical stakeholders can work on the "Interoperability Stack," creating the "common technical language" needed for national frameworks to "talk" to one another. This prevents "trustworthy" status in one region from becoming a security liability in another. The "Compute Prerequisite" Lab: A dedicated track for developing nations to access high-performance computing, transforming capacity-building from a "theoretical exercise" into a practical, "security necessity".Evidence-Based "Hype Audits": Using the Scientific Panel's benchmarks, participants can audit current AI tools to distinguish between marketing hype and verifiable system capabilities. Open-Source Sandbox: A collaborative space where participants can audit and adapt open-source models, fostering transparency as the "ultimate security audit" and preventing "security dependencies".By structuring the dialogue around these "Day One" practical outputs, the forum ensures that "safe and secure" is treated as a verifiable state rather than a vague policy goal.

To move from high-level ethics toward a harmonized, technical framework, the following approaches offer concrete solutions for effective AI governance:" Secure-by-Design" Architecture: Implementing AI as a complex stack of compute, data, and models that requires security to be the foundation rather than a "bolt-on" feature. Global AI "Red-Teaming" Network: A collaborative platform designed for the real-time sharing of vulnerability data and threat models for Large Language Models (LLMs).The "Interoperability Stack": A set of technical standards that enable disparate national frameworks to "talk" to one another, ensuring consistency across borders. Technical Telemetry and Benchmarking: Utilizing an Independent Panel to deliver benchmarks that distinguish between "marketing hype" and verifiable system capabilities. Open-Source as a Security Audit: Promoting open-source models to allow developing nations to audit and secure tools within their own contexts, thereby preventing "security dependencies". Adversarial Resilience Practices: Shifting from philosophical safety to measurable resilience against technical attacks, such as data poisoning and model inversion. Compute-Centric Capacity Building: Treating access to high-performance computing as a security necessity rather than just an economic issue to bridge the digital divide. Robust Human Oversight: Clustering governance discussions around verifiable mechanisms for compliance with international law. These mechanisms ensure that "safe and secure" becomes a verifiable state rather than an abstract goal.