Concordia AI

First, reaffirming support for the Panel to provide international guidance in the fast-moving AI landscape — for example, on governance of agentic systems that can increasingly act autonomously, use tools, and chain decisions without human intervention. These systems are already deployed across critical sectors without agreed safety assessments. The Dialogue should task the Panel with producing accessible scientific recommendations on such developments ahead of the 2027 meeting. Second, a shared commitment to pursue AI safety as a global public good. The Dialogue could agree on principles for voluntary, good-faith cooperation, such as exploring ways to share risk management practices, evaluation standards, and model provenance information over time, while recognizing that capacity remains concentrated. Third, an agreed roadmap to close the safety divide, not only the innovation divide. The first meeting can commission work on infrastructure such as incident response templates, shared evaluation benchmarks, and preparedness planning — to be developed between sessions and reviewed in New York in 2027.

• Safe, secure and trustworthy AI
• Interoperability of governance approaches
• Transparency, accountability, and human oversight
• Open-source software, open data and open AI models

Active engagement is needed to provide support for the Panel to incorporate latest research on safe, secure and trustworthy AI. Given that tools-using agents are already outrunning existing safety assumptions, without an explicit directive to stay current with these rapid developments, the Panel risks producing consensus on past problems. Tasking it now to deliver accessible recommendations by 2027 on safety and transparency ensures the second meeting has a shared scientific basis for action. Equally important, open-weight and open-source models are proliferating faster than the governance capacity to evaluate them locally.The Dialogue should task the Panel with identifying gaps in the scientific infrastructure needed to govern open-weight and open-source models, such as methods for training data curation, documenting model provenance, and localised evaluation benchmarks - and with proposing cooperative steps to build distributed evaluation capacity, consistent with Resolution 79/325.

Domain AI-enabled risks in biosecurity and cybersecurity. As frontier AI models grow more capable, they increasingly interact with domains where misuse carries severe consequences. In biosecurity, advanced models can compress the expertise previously required to access technical knowledge relevant to pathogen engineering. In cybersecurity, AI-enabled tools are expanding the scale and sophistication of offensive operations available to a wider range of actors. These are not speculative concerns - frontier AI developers are already conducting internal evaluations for dangerous capability thresholds in both domains. What is missing is an international framework for how such evaluations are conducted, shared, and acted upon.

Our work facilitating AI safety dialogues across Asia gives us direct visibility into governance gaps that are compounding rapidly. China's open-weight and open-source AI ecosystem illustrates the challenge precisely. Models like DeepSeek R1 and the Qwen series are being adopted at extraordinary speed across sectors and jurisdictions. In particular, open-source agentic systems like OpenClaw are increasingly combining agentic decision-making with real-world consequences. The governance implications are significant: these models are being integrated into critical applications globally, yet there are no shared evaluation standards, no comparable provenance documentation requirements, and no internationally recognized safety benchmarks applied at the point of deployment. The speed of adoption consistently outpaces the development of governance frameworks, domestic and international. This is just one example of a governance gap that the Global Dialogue is uniquely positioned to address.

The Dialogue's unique value is convening authority across all Member States as a truly inclusive body. We would like to call attention to three key functions it can serve: First, incorporate recommendations of the Independent International Scientific Panel on Artificial Intelligence in governance proposals. The Panel needs to be established as a consensus-building mechanism that builds on existing work, such as the International AI Safety Report, to deliver continuous, technically authoritative assessments of AI benefits and risks that all Member States can act upon. Second, establish a global safety infrastructure sharing mechanism: a structured process for distributing evaluation standards, red-teaming methodologies, incident reporting frameworks, and model assessment tools to states that currently lack capacity to develop them independently. Third, create interoperability bridges between existing governance initiatives: the AI Safety Summits, EU Code of Practice, regional frameworks in order for distributed progress to become cumulative rather than fragmented.

The Dialogue should build directly on existing technical foundations , including the International AI Safety Report, national AI Safety Institutes and the International Network for Advanced AI Measurement, Evaluation and Science as well as the AI Safety Summits (Bletchley Park, Seoul, Paris, India AI Impact Summit). These initiatives have produced research priorities, risk assessment and governance frameworks that remain fragmented and limited in reach. The Dialogue should scale these models multilaterally by connecting technical capacity with policy infrastructure across all regions. Its unique added value is not originating new work but giving these initiatives universal reach through the only platform where all Member States participate.

Technical and scientific communities, including AI Safety Institutes should contribute through open scientific exchanges: technically grounded, inclusive across governance traditions, and feeding directly into the first Global Dialogue. Civil society and affected communities should have formal input mechanisms beyond mere observer status, particularly on social, economic, and developmental implications that technical assessments alone cannot capture. Industry stakeholders should contribute through structured disclosure: sharing evaluation methodologies, incident data, and capability assessments under clear terms that serve the Panel's independent analysis rather than corporate positioning. Most importantly, the Dialogue's format should ensure that each convening produces specific outputs: technical assessments, governance recommendations, capacity-sharing commitments. Sessions should be organized around the Panel's findings, with dedicated tracks for discussing the distribution of benefits, frontier risk assessment, agentic AI oversight etc., ensuring that scientific consensus directly informs multilateral action.