Source-backed entity profiles

Core Chinese internet-governance regulator for AI-related online information services. Primary-source materials show CAC co-issuing the 2023 Interim Measures for Generative AI Services, the 2025 AI-generated content labeling measures, the 2022 deep-synthesis rules, and the 2024 national AI standardization guide.

Central industry and telecom ministry with a major role in the China AI standards and evaluation stack. Primary sources show MIIT co-issuing CAC-facing AI service rules and the 2024 national AI standardization guide, and serving as the home ministry for affiliated institutions including CAICT and CESI.

China's main national cybersecurity standardization committee with an increasingly visible AI-security and AI-governance role. Official sources show AI-specific standards and guidance on generative AI service security, content labeling methods, AI safety governance, AI application ethics, and the 2026 formation of WG9 for AI security standards.

National information-technology AI subcommittee under the broader TC28 standards structure. Official sources show CESI as the secretariat unit and confirm the committee's role in AI standardization work.

MIIT-affiliated research institution that is central to China's applied AI governance, benchmarking, and industry-self-regulation stack. Official sources show CAICT's direct MIIT affiliation, its secretary-general-unit role in jointly founding AIIA, its role in launching the AI Risk Management System and AI Safety Benchmark through AIIA, and its Fangsheng benchmarking system with a dedicated safety testing dimension.

Needed because multiple official CAICT sources treat AIIA as the umbrella platform through which CAICT launched safety-governance structures, AI safety commitments, and AI safety benchmarks.

Best-evidenced through official CAICT reports rather than a dedicated committee page. Official CAICT sources state that CAICT, relying on AIIA, prepared a Safety Governance Committee in late 2023, released an AI Risk Management System, and later linked AIIA to AI safety commitments and the AI Safety Benchmark.

MIIT-affiliated standards institute with a central role in China's AI standards and model-evaluation infrastructure. Official sources place CESI as secretariat host for TC28/SC42 and show CESI releasing both a general large-model evaluation standard and the Qiúsuǒ national-standard benchmark system.

Official site evidence supports CSTC as an AI assessment actor rather than a pure standards drafter. The homepage lists AI evaluation and AIM AI management maturity assessment services, and the center's Sadie Lab activity stream includes an AI large-model safety-governance seminar and AI safety workgroup launch.

Major Shanghai AI research institution with visible roles in evaluation infrastructure, AI security standardization, and PR2G-imported frontier-risk framework work with Concordia AI.

Officially evidenced through a government-domain repost from Cyberspace Shanghai. The retrieved source states the lab was jointly established by SHLAB's Governance Research Center and the Shanghai Information Security Testing and Certification Center and unveiled at WAIC 2024.

Official SHLAB reporting attributes WG9's formal establishment to a TC260 notice and identifies SHLAB's director as group head, making WG9 a material node in the China AI security standards stack.

Open evaluation infrastructure from Shanghai AI Lab that explicitly includes safety as a model capability dimension.

Non-profit new-type AI R&D institution that is significant in China's model-evaluation and AI-safety dialogue ecosystem. Official sources show BAAI maintaining FlagEval as an open evaluation platform and hosting the Beijing AI Safety International Dialogue that produced the Beijing AI Safety Consensus.

BAAI’s public evaluation platform for large and foundation models, including safety and values dimensions.

Beijing's dedicated AI safety-and-governance institute appears to have evolved from a 2024 laboratory phase into a formally established institute in April 2025. Official sources also show a department structure that explicitly distinguishes AI Safety from AI Security and includes guardrails, ethics, and risk-response work.

Official Chinese institutional coverage positions CnAISDA primarily as a China-side dialogue network corresponding to an AI safety institute for international engagement, not automatically as a domestic frontier-model testing regulator. Tsinghua and SQZ sources explicitly identify CAICT, BAAI, and SHLAB as member units.

The official committee homepage shows a distinct ministry-level AI standardization architecture with dedicated working groups and a CAICT-linked operational base, making it highly relevant to the China standards stack even though it was not in the starting cluster.

Strong public technical-evaluation node with explicit guides for LLM-system safety evaluation and red teaming.

Major Korean public evaluation node with explicit risk categories, benchmark infrastructure, and public model-evaluation activity.

Publicly structured as a designation-plus-partnership model rather than a standalone statutory institute: DTC is the designated AISI and IMDA is the overall government agency.

Singapore’s strongest public technical-testing ecosystem node, relevant here because it operates Project Moonshot for LLM benchmarking and red teaming.

ASEAN regional working group on AI governance. The Bangkok Digital Declaration directly supports a coordination remit that includes AI policy, governance, safety, laws, regulations, and standards, but does not establish supranational regulatory authority.

Social enterprise with offices in Beijing and Singapore. PR2G treats Concordia as a highly APAC-linked comparator across China/Singapore research, standards and policy engagement, and frontier-risk framework work with Shanghai AI Laboratory.

Non-profit dialogue convenor whose flagship series is IDAIS. PR2G imports it as directly APAC-relevant through Beijing and London IDAIS dialogue architecture.

Recurring international AI safety dialogue series. PR2G strengthens the record with normalized SAIF/IDAIS evidence and imports Beijing/London session events where source IDs are available.

Tsinghua-linked AI international governance research node included for policy, governance, and cross-border dialogue mapping.

Tsinghua-linked center that appears in public records for U.S.-China and China-EU AI and international security dialogues. PR2C keeps these descriptions high-level and models the rows as public dialogue channels, not operational security relationships.

Host agency in which Japan AISI is placed and where the AISI secretariat was set up.

Cabinet Office-linked inter-ministerial coordination mechanism around Japan AISI, with ministries, research institutes, IPA, and AISI represented.

The ministry that launched Korea AISI and publicly ties it to post-Seoul-Summit implementation and international AI safety networking.

The clearest public institutional home for Korea AISI in reviewed sources.

Consortium formed at Korea AISI’s launch to support framework work, subgroup collaboration, and joint-testing discussions.

NTU-based centre designated as Singapore AISI and serving as the technical backbone of the Singapore model.

Update IMDA with Singapore’s assurance and testing implementation roles, including A.I. Verify launch, the Global AI Assurance Sandbox with AI Verify Foundation, the Singapore AI Safety Red Teaming Challenge, and 2026 testing-standardisation activity.

Core assurance framework in Singapore’s public stack; should be separated from Singapore AISI itself.

Open-source LLM evaluation toolkit that combines benchmarking and red teaming for LLMs and LLM applications.

Singapore-linked assurance programme for deployed or deploying GenAI applications, represented as application-level assurance and market-building infrastructure rather than foundation-model safety testing.

Distinct third-party testing accreditation programme in Singapore’s assurance ecosystem.

Public co-host institution for repaired China-EU AI and International Security Dialogue round records. The atlas models this as a dialogue-convening role only.

China-side research institute visible in this PR2C import through the source-backed CnAISDA Paris side-event record. Broader IDAIS-related claims are held until exact URLs are repaired.

Chinese frontier-model developer with public disclosures centered on a transparency page plus open-weight DeepSeek-V3 repository and license documentation. This backfill does not identify a public preparedness policy, dangerous-capability evaluation framework, or external frontier-safety audit.

Alibaba Qwen model-family record backed in this pass by official repository and model-card documentation. The public evidence supports open-weight/model-card disclosures, not a public dangerous-capability or release-gating framework.

Chinese frontier-model developer whose BigModel documentation surfaces API access, content-safety controls, model-evaluation tooling, GLM-4.5 model information, and privacy policy material. These are platform and evaluation disclosures, not a public preparedness framework.

Baidu ERNIE/Qianfan record backed by Qianfan platform, user-agreement, intervention, API, and content-marking documentation. Public disclosures are strongest on platform governance, content controls, data handling, and generated-content labeling.

Tencent Hunyuan record backed by Tencent Cloud product, API, compatibility, moderation, and image-labeling documentation. The evidence supports platform controls and alignment-language disclosures, not a public frontier-risk governance policy.

ByteDance/Doubao record backed by Volcengine Ark and Doubao service documentation. Public materials emphasize platform terms, API access, and model-evaluation workflow disclosures; snippet-only platform-scope claims were held out of evidence and relationships.

Moonshot AI Kimi record backed by Kimi privacy, service-agreement, platform, and Kimi K2 model-card materials. Public evidence supports privacy, developer data-responsibility, tooling, and model-card disclosures rather than release gates or advanced-risk governance.

MiniMax record backed by about, API, privacy, model-page, and content-labeling materials. Public disclosures are mostly API/platform, privacy, benchmark, and content-provenance oriented; snippet-only API-key and labeling evidence notes were held.

StepFun record backed by privacy, user-agreement, platform, migration, and reasoning API pages. Because PR2D evidence for detailed API and agent-framework claims was snippet-only, this import keeps StepFun claims at the general platform-disclosure level.

01.AI/Yi record backed by repository, model-card, and licensing source rows. This pass treats Yi as primarily an open-weight documentation case and does not import snippet-only claims about training details, context windows, or license language.

Official Zhipu platform surface used in PR2D for source-backed API access, content-safety, and model-evaluation disclosures.

Official Baidu Qianfan platform surface used in PR2D for model services, intervention rules, API access, and content-marking controls.

Official Moonshot AI Kimi API platform surface used in PR2D for source-backed tooling, privacy, service-agreement, and model-card disclosures.

Regional intergovernmental organization represented here only for source-backed ASEAN AI governance coordination: the 2024 voluntary AI governance guide, the 2025 generative-AI expansion, the Bangkok Digital Declaration, and a Malaysia official ASEAN AI Safe announcement. Held roadmap and search-snippet claims are not treated as operational evidence.

Malaysian government-linked implementation and coordination node in this source pack, connected to NAIO incubation, AI action-plan procurement, and a Malaysia official announcement about ASEAN AI Safe. The public sources do not establish MyDIGITAL as an AI safety testing body.

Thai digital-governance agency represented here through official AI governance guidance, AI Governance Center activity, ethics principles, and draft AI-law consultation sources. The imported rows do not treat the draft-law consultation as enacted law.

Philippine government research institute represented here as a public AI R&D implementation actor, including DOST-ASTI AI programme materials, SkAI-Pinas-linked work, and the launch of NAICRI.

BRIN research organization represented with limited public evidence that it hosts an AI and cybersecurity research center page. Details on mandate, staffing, and programme scope remain sparse in this source pack.

Regional AI safety network initiative supported here by a Malaysia official announcement. The atlas does not treat it as an operational testing network until ASEAN publishes public implementation details.

IMDA-organized multicultural and multilingual GenAI red-teaming challenge focused on harmful bias stereotypes across Asian contexts.

Malaysia national AI coordination node launched in December 2024. Sources describe policy and regulatory framework shaping, innovation, adoption, and talent functions, but do not establish statutory enforcement powers.

Set of NAIO specialised working groups publicly announced to drive Malaysia’s AI agenda with technology, academic, industry, government, and civil-society participation.

ETDA-hosted AI governance center framed as a cross-network cooperation platform for governance framework development, consultation, knowledge exchange, and networking.

Official national AI strategy and action-plan programme page for Thailand, including readiness work on ethics, law and regulation and a strategy line for AI sandboxes.

BRIN page indicating a dedicated AI and cybersecurity research center under the Electronics and Informatics research organization. Public detail in this source pack is limited.

Philippine national institutional anchor for AI research, advanced computing, and innovation launched by DOST-ASTI in February 2026.

Philippine AI R&D programme represented because DOST-ASTI’s official AI page identifies ASTI-ALaM as part of the broader SkAI-Pinas umbrella.

Public PKU team focused on LLM alignment, Safe-RLHF, multimodal alignment, and alignment-brittleness research.

BAAI-led public safety platform spanning red teaming, defense, interpretability, deception, and alignment.

Publicly described as working on large-model interpretability and safety alignment.

BIGAI is included for TongTest and explicit ability-plus-value-alignment framing; the row remains medium-confidence because the work is broader AGI evaluation rather than pure frontier LLM safety.

Public HKUST materials show direct LLM security relevance and broader trustworthy-LLM framing around security and reliability risks.

Publicly visible through HKGAI-V1, which is framed around Hong Kong-specific alignment, safety, and contextual suitability.

Key Japanese public node for technical LLM safety, with a Safety WG and safety symposiums.

Public safety working group inside Japan’s open LLM ecosystem, visible through NII and LLM-jp resources.

Corporate technical-safety team with public Japanese benchmarks for safety-boundary testing and guardrails.

Relevant mainly through public interpretability and trustworthy-foundation framing plus adjacent Japanese LLM safety dataset work. This row remains medium-confidence and is not labeled as a core frontier AI safety lab.

University-based Korean group with direct public evidence of large-model security and safety benchmarking using prompt-injection and jailbreak attacks.

KAIST lab with direct public work on jailbreak defense for misaligned instructions and transformer interpretability.

Publicly relevant through agent-safety, backdoor, and LLM-safety work such as AgentSpec and BackdoorLLM.

NUS group relevant through formal methods with LLM agents, trusted AI, and the TrustLLM benchmark.

University parent institution for source-backed HKUST technical safety and LLM-security rows in PR2F, including Shuai Wang group materials and HKGAI-V1 public activity.

NIST-hosted U.S. comparator entity for AI standards, testing, collaborative research, voluntary agreements, and broad risk-evaluation work. PR2G treats CAISI as an institutional comparator and does not import operational benchmark details.

State-backed advanced-AI safety and security institute used as a Western comparator for evaluation, mitigation testing, and international coordination, with APAC pathways through Australia and India-linked coalition activity.

European Commission office implementing the AI Act, especially for general-purpose AI. It is modeled as a regulatory and evaluation comparator rather than a lab-style AISI.

OECD-operated AI policy observatory hosting the AI Principles, AI Incidents and Hazards Monitor, Hiroshima AI Reporting Framework, and OECD-GPAI materials. It is the strongest standards/reporting comparator in PR2G.

U.S. federal standards institution imported only as CAISI’s parent/host and the AI Consortium source for broader AI metrology and standards context.

Current official UK government naming found in PR2G for an international AISI-style measurement, evaluation, and science network. Participant list is not treated as complete.

APAC counterpart imported because the PR2G sources directly tie it to UK AISI through a formal 2026 MoU and UK Alignment Project coalition materials.

Oxford governance initiative imported only for its directly supported IDAIS role, especially the London 2026 co-hosting claim and Beijing/series linkage.