AIAI Briefing — December 15, 2025
The International Network of AI Safety Institutes expands coordination efforts as member nations align on model evaluation protocols, red-team testing standards, and incident reporting frameworks ahead of 2026 regulatory milestones.
Executive briefing: On , the International Network of AI Safety Institutes (INASI) released coordinated guidance on frontier AI model evaluations, red-team testing protocols, and cross-border incident reporting. This framework represents the culmination of 18 months of multilateral negotiations following the Bletchley Declaration and builds on bilateral agreements between the UK AI Safety Institute, US AI Safety Institute, and counterparts in the EU, Japan, Canada, and Singapore. Organizations developing or deploying advanced AI systems should prepare for harmonized evaluation requirements entering force in 2026.
Core Framework Elements
The INASI guidance establishes four pillars for coordinated AI safety governance:
Model Evaluation Protocols: Standardized evaluation methodologies for frontier AI systems, including capability assessments, dangerous capability identification, and misuse potential analysis. The protocols define thresholds triggering enhanced scrutiny, drawing on research from AI safety labs and academic institutions. Evaluations must assess capabilities across domains including biological and chemical weapons synthesis, cyber offense, manipulation and deception, and self-replication.
Red-Team Testing Standards: Harmonized requirements for adversarial testing before model deployment, including structured attack taxonomies, documentation templates, and minimum testing coverage. Organizations must demonstrate red-team findings have informed deployment decisions and safety mitigations. The standards establish competency requirements for red-team practitioners and encourage third-party verification.
Incident Reporting Framework: Common definitions for AI safety incidents, severity classification schemas, and reporting timelines. Member institutes commit to sharing incident information within 72 hours for high-severity events and maintaining a confidential incident database for trend analysis. The framework distinguishes between technical failures, misuse events, and near-misses requiring investigation.
Information Sharing Protocols: Mechanisms for secure exchange of evaluation results, vulnerability disclosures, and research findings among member institutes. The protocols balance transparency with confidentiality concerns, establishing tiered disclosure with appropriate protections for sensitive capability information.
Key Requirements for AI Developers
Organizations developing frontier AI systems should prepare for the following obligations under INASI member jurisdictions:
Pre-deployment Evaluations: Models exceeding defined compute thresholds (generally aligned with EU AI Act systemic risk criteria) must undergo structured evaluations before deployment. Developers must provide technical documentation, access for external evaluators, and evidence of safety mitigations. Evaluation timelines range from 30 to 90 days depending on model capabilities and risk profile.
Ongoing Monitoring: Post-deployment monitoring requirements include capability tracking, misuse detection, and incident documentation. Developers must establish monitoring pipelines capable of detecting emergent behaviors, coordinated misuse campaigns, and performance degradation affecting safety properties.
Transparency Reporting: Annual transparency reports covering model capabilities, safety evaluations, red-team findings, and incident summaries. Reports must follow standardized templates enabling cross-jurisdictional comparison and aggregated analysis. Certain information may be provided confidentially to regulators rather than published publicly.
Cooperation Obligations: Requirements to cooperate with safety institute inquiries, provide requested information within specified timelines, and participate in coordinated investigations. Developers must designate regulatory contacts and establish secure communication channels with relevant authorities.
Implications for Organizations
The INASI framework has significant implications for organizations across the AI value chain:
Frontier AI Developers: Companies developing advanced AI systems face compliance obligations across multiple jurisdictions. Early engagement with safety institutes, investment in evaluation infrastructure, and establishment of safety governance functions are essential. Developers should allocate resources for pre-deployment evaluation periods and build relationships with external evaluators.
AI System Deployers: Organizations deploying third-party AI systems should conduct due diligence on supplier evaluation practices and obtain documentation of safety assessments. Deployment decisions should incorporate evaluation findings and establish monitoring requirements proportionate to system risks.
Research Institutions: Academic and non-profit research organizations contribute to evaluation methodology development and may participate in external evaluations. Institutions should engage with safety institutes on research priorities and consider how their work supports the broader evaluation ecosystem.
Critical Infrastructure Operators: Organizations operating critical infrastructure should assess AI deployment risks in light of INASI incident definitions and reporting requirements. Integration of AI governance with existing operational resilience frameworks ensures coordinated response to AI-related incidents.
Recommended Actions
Organizations should take the following steps to prepare for INASI-aligned requirements:
Immediate (0-3 months): Conduct inventory of AI systems to identify those potentially subject to enhanced evaluation requirements. Assess current evaluation practices against INASI standards and identify gaps. Brief leadership on coordination developments and resource implications.
Near-term (3-6 months): Establish relationships with relevant national safety institutes and participate in consultations on implementation guidance. Develop or enhance red-team testing programs aligned with emerging standards. Build documentation pipelines for transparency reporting.
Medium-term (6-12 months): Implement monitoring infrastructure meeting post-deployment requirements. Train staff on incident classification and reporting procedures. Integrate AI safety governance with existing compliance frameworks for coordinated oversight.
Ongoing: Monitor guidance updates from safety institutes and participate in industry coordination on implementation challenges. Track evaluation methodology developments and incorporate learnings into internal practices. Maintain dialogue with regulators on emerging risks and compliance approaches.
Coordination Mechanisms and Governance
The INASI operates through several coordination mechanisms:
Plenary Assembly: Annual meeting of all member institutes to review progress, adopt guidance updates, and set strategic priorities. Observer nations may participate in discussions but do not vote on binding decisions.
Technical Working Groups: Subject-matter groups addressing evaluation methodologies, red-team standards, incident analysis, and emerging risks. Working groups produce draft guidance for Plenary Assembly consideration and coordinate joint research initiatives.
Secretariat Functions: Administrative support including meeting coordination, document management, and communications. The secretariat rotates among member institutes on a two-year cycle.
Joint Evaluation Pool: Shared roster of qualified evaluators available for cross-border assessments. Pool members undergo common training and certification, enabling efficient deployment for time-sensitive evaluations.
Challenges and Open Questions
Several challenges remain as the network matures:
Capability Thresholds: Defining compute and capability thresholds triggering evaluation requirements remains contentious. Thresholds must balance capturing genuinely risky systems against creating excessive compliance burdens for beneficial applications. The network commits to regular threshold reviews as technical understanding evolves.
Confidentiality Tensions: Balancing transparency with protection of sensitive capability information presents ongoing challenges. Information sharing protocols must prevent disclosure of dangerous capabilities while enabling meaningful public accountability.
Non-Member Jurisdictions: Coordination gaps persist where major AI development occurs outside member jurisdictions. The network seeks to expand membership and establish informal cooperation arrangements with non-members while avoiding legitimizing inadequate safety practices.
Resource Constraints: Safety institutes face capacity limitations relative to the pace of AI development. Prioritization of evaluation targets, efficient use of shared resources, and investment in evaluation tooling help address constraints but do not fully resolve them.
Zeph Tech Analysis
The INASI framework represents a significant maturation of international AI governance from aspiration to operational coordination. While challenges remain, the establishment of common evaluation standards, incident reporting protocols, and information sharing mechanisms provides foundations for effective oversight of frontier AI systems.
Organizations should view INASI compliance not merely as a regulatory obligation but as an opportunity to demonstrate commitment to responsible AI development. Early engagement with safety institutes, investment in evaluation capabilities, and transparent reporting can build trust with regulators and stakeholders while informing product safety decisions.
The framework's evolution will depend on continued political commitment, technical progress in evaluation methodologies, and industry cooperation. Organizations that contribute constructively to implementation challenges will help shape practical requirements while positioning themselves favorably for regulatory interactions.
Zeph Tech will continue monitoring INASI developments and providing guidance on compliance approaches as implementation progresses through 2026.
Continue in the AI pillar
Return to the hub for curated research and deep-dive guides.
Latest guides
-
AI Workforce Enablement and Safeguards Guide — Zeph Tech
Equip employees for AI adoption with skills pathways, worker protections, and transparency controls aligned to U.S. Department of Labor principles, ISO/IEC 42001, and EU AI Act…
-
AI Incident Response and Resilience Guide — Zeph Tech
Coordinate AI-specific detection, escalation, and regulatory reporting that satisfy EU AI Act serious incident rules, OMB M-24-10 Section 7, and CIRCIA preparation.
-
AI Model Evaluation Operations Guide — Zeph Tech
Build traceable AI evaluation programmes that satisfy EU AI Act Annex VIII controls, OMB M-24-10 Appendix C evidence, and AISIC benchmarking requirements.
Comments
Community
We publish only high-quality, respectful contributions. Every submission is reviewed for clarity, sourcing, and safety before it appears here.
No approved comments yet. Add the first perspective.