Written and maintained by CASRAI Editorial Board
Last updated
Search for “network of ai safety institutes” and you’ll find two different things layered on top of each other: a coordination body among national AI safety/security institutes that was announced in 2024, and a renamed, refocused version of that same body that exists today under a different name entirely. This guide is about that coordination body — not about any single national institute’s own testing program, which CASRAI covers separately for CAISI and the UK AI Security Institute, and not about the corporate commitments made at the Seoul AI Safety Summit, which get their own guide. Verified against the UK AI Security Institute’s own published material, September 2026.
What the Network is
The International Network of AI Safety Institutes was established in November 2024 as a coordination body bringing together national government institutes responsible for evaluating advanced AI systems — the same kind of body as the U.S.’s CAISI and the UK’s AI Security Institute, which both renamed themselves in 2025, but sitting above them rather than being one of them. Ten members currently participate: Australia, Canada, the European Union, France, Japan, Kenya, the Republic of Korea, Singapore, the United Kingdom, and the United States.
The Network does not test models itself and has no authority over any member institute’s own evaluation program. What it coordinates is methodology: how member institutes define evaluation objectives, select benchmarks, and report results, so that one country’s evaluation findings are comparable to another’s rather than each institute inventing its own incompatible approach.
The rename to NAAIMES
The Network’s current formal name is the International Network for Advanced AI Measurement, Evaluation and Science (NAAIMES) — UK AISI’s own published material describes this as “a refocusing of the former International Network of AI Safety Institutes.” As with CAISI and UK AISI’s own name changes, the rename tracks a narrowing of stated mission: from a general “AI safety” framing toward a more specific focus on the measurement science underlying evaluation — benchmark validity, reproducibility, and comparability — rather than safety policy more broadly. The ten-member list did not change with the rename; what changed is the name and the stated emphasis, the same pattern already documented for the individual national institutes in this cluster.
What the Network has actually produced
NAAIMES’s output so far is best-practice documentation and convening, not joint testing of specific models. Member institutes have published a first best-practice guidance document on evaluation methodology, covering how to define evaluation objectives and select appropriate benchmarks, how to ensure comparability between different institutes’ evaluations, how to iterate on capability elicitation (the techniques used to draw out a model’s maximum capability during testing), and how evaluations should be conducted and their results tracked.
Members have also convened in person: alongside the NeurIPS conference in San Diego in 2025 for technical workshops with outside researchers and industry, and on the margins of the 2026 International Conference on Machine Learning in Seoul to continue building consensus on evaluation science. The Network operates a rotating Network Coordinator role among its members; the UK holds that role in 2026 and has said it will use the position to turn the group’s shared learning into more detailed best-practice documentation. A further meeting is planned at the India AI Impact Summit, where members intend to compare lessons learned and test their agreed approaches against real-world use cases.
How this differs from bilateral arrangements between individual institutes
The Network is easy to confuse with the separate, older bilateral relationship between CAISI and UK AISI specifically — the April 2024 agreement to test models jointly, which predates the Network by seven months and predates both institutes’ 2025 renames by close to a year. That bilateral arrangement is a working relationship between two institutes’ testing programs; the Network is a ten-member forum for evaluation-methodology consensus that neither tests models jointly on its members’ behalf nor replaces any individual institute’s own bilateral agreements with AI developers. An institute can belong to the Network and still run its own separate pre-deployment testing relationships with frontier labs, exactly as CAISI and UK AISI do.
Where this connects to NIKOLAI
CASRAI’s NIKOLAI dictionary defines several of the exact terms NAAIMES’s own best-practice guidance is trying to standardize. NIKOLAI’s N5 · Evidence and Evaluations track includes an elicitation method element — defined as “the techniques and conditions used to draw out a model’s maximum capability during an evaluation, recorded as a property of the evaluation run, together with the declared interpretation of the result” — which is the same concept NAAIMES’s guidance addresses under “iterating on capability elicitation.” NIKOLAI’s N8 · Transparency and Review track separately defines evaluator independence, relevant to any comparison of how rigorously a given national institute’s evaluations are insulated from the developer being tested. Neither NIKOLAI element is endorsed by NAAIMES or any of its ten member institutes; they’re CASRAI’s own vocabulary, offered as a stable reference point for readers mapping one institute’s terminology against another’s.
FAQ
Which countries belong to the International Network of AI Safety Institutes / NAAIMES?
Ten: Australia, Canada, the European Union, France, Japan, Kenya, the Republic of Korea, Singapore, the United Kingdom, and the United States.
Is this the same as the CAISI–UK AISI joint testing agreement?
No. The CAISI–UK AISI bilateral testing relationship dates to April 2024 and is a two-institute arrangement to test models together and share findings. The Network (now NAAIMES) is a separate, ten-member forum focused on evaluation-methodology consensus, established seven months later in November 2024. Membership in one doesn’t require or replace the other.
Did the Network test models jointly?
Not as an institution. NAAIMES’s published output is best-practice guidance and convening among member institutes’ own evaluation teams, not a joint evaluation program that tests specific models on members’ behalf.
Why was the Network renamed to NAAIMES?
UK AISI’s own material describes the new name as “a refocusing” of the original International Network of AI Safety Institutes, narrowing the stated mission toward measurement science — evaluation methodology, comparability, and reproducibility — rather than AI safety policy more broadly. The same ten members continued under the new name.
Who runs the Network?
It operates a rotating Network Coordinator role among its ten members. The UK holds the role in 2026.
