Ethics bodies and open models, but no framework for dangerous capabilities
Overall rating “limited”. No deduction for controversies. The superscript numbers link directly to the source.
1 · Responsibility
The ESG report documents an ethics review commission and board responsibility, but independent reviewers find neither a published framework nor a whistleblowing channel.
What is documented
- The 2025 ESG report assigns ESG responsibility to the board and names four committees, including an ESG and strategy committee1
- An ethics review commission covers model development from planning to decommissioning; there are also rules on algorithm safety and model retirement1
- The company states certificates under ISO/IEC 42001, 27001 and 27701 without naming the certification bodies1
- In Seoul in 2024, Zhipu signed the Frontier AI Safety Commitments as the only Chinese firm5 11
- Z.ai answered the survey for the FLI AI Safety Index 20262
The case against
Commitments and oversight
- According to the FLI, the safety framework with risk thresholds committed to in Seoul had not been published by June 20265 4 2
- The FLI grades Governance & Accountability F and finds no clear governance structure and no whistleblowing channel2
- The composition and powers of the ethics review commission and the certification bodies are not disclosed1
Sources for Responsibility (12)
- Knowledge Atlas Technology JSC (Zhipu): 北京智譜華章科技股份有限公司 2025年環境、社會及管治報告 (first ESG report, reporting year 2025) · 3 November 2026 · primary source
- Future of Life Institute: AI Safety Index – Summer 2026 (Digital, 14 Jul) · 14 July 2026 · primary source
- HKEX / Knowledge Atlas Technology JSC: Announcement – Appointment of Overall Coordinators (Rule 12.01C) · 19 December 2025 · primary source
- Future of Life Institute: AI Safety Index – Summer 2025 · 17 July 2025 · primary source
- UK Government (DSIT): Frontier AI Safety Commitments, AI Seoul Summit 2024 · 21 May 2024 · primary source
- NYU Shanghai RITS (Utku Ege Tuluk): GLM-5.3: Z.ai Scales Post-Training, Holds the Weights Back · 17 August 2026 · secondary source
- Mervin Praison (mer.vin): GLM-5.3 Ships Open Weights After Two Weeks as an API-Only Model · 28 August 2026 · secondary source
- 新京报 (Beijing News), 罗亦丹: 智谱GLM-5.3发布,涌现网络安全能力 · 14 August 2026 · secondary source
- Z.ai: Z.ai Terms of Use · 14 April 2026 · secondary source
- U.S. Department of Commerce, Bureau of Industry and Security (reproduced from the Federal Register via FR Tracker): Addition of Entities to and Revision of Entry on the Entity List, 90 FR 4617 (Doc. 2025-00704) · 16 January 2025 · primary source
- Carnegie Endowment for International Peace (Scott Singer): DeepSeek and Other Chinese Firms Converge with Western Companies on AI Promises · 28 January 2025 · secondary source
- The Star (Reuters, Joe Cash): Chinese AI-related firm Zhipu says strongly opposes inclusion in US export control entity list · 16 January 2025 · secondary source
2 · Safety
CAISI found GLM-5.2 robust against agent hijacking, but it refused no exploit task, and GLM-5.3 is the most cyber-capable open model.
What is documented
- CAISI found no successful agent hijacking of GLM-5.2; the model was more robust than other open models but less robust than leading US models4
- According to the ESG report, encryption, multi-factor authentication, red-team tests and emergency drills protect the systems; by its own account there were no data breaches in 20251
- The FLI acknowledges some meaningful incident-response infrastructure2
- According to reports, Z.ai held back the weights of GLM-5.3 for two weeks to complete safety evaluation and hardening8 9 7
- The terms of use set out prohibited uses, suspension rights and a reporting channel11
The case against
Dangerous capabilities
- CAISI rates GLM-5.3 as the most cyber-capable open-weight model to date; the weights have been public since 28 August 20266 8
- GLM-5.2 refused none of the ten agentic exploit tasks and answered sensitive biology questions more often and in more detail than US models4
- The FLI grades risk assessment F; Z.ai's own evaluations of dangerous capabilities are not published2
Sources for Safety (13)
- Knowledge Atlas Technology JSC (Zhipu): 北京智譜華章科技股份有限公司 2025年環境、社會及管治報告 (first ESG report, reporting year 2025) · 3 November 2026 · primary source
- Future of Life Institute: AI Safety Index – Summer 2026 (Digital, 14 Jul) · 14 July 2026 · primary source
- GLM-4.5 Team, Zhipu AI & Tsinghua University (arXiv): GLM-4.5: Agentic, Reasoning, and Coding (ARC) Foundation Models (arXiv:2508.06471v1) · 8 August 2025 · primary source
- NIST, Center for AI Standards and Innovation (CAISI): Assessment of Z.ai's GLM-5.2 · 8 July 2026 · primary source
- SPLX (Zscaler), Mateja Vuradin: GLM-4.5. Can It Pass the Enterprise AI Security Test Where Kimi K2 Failed? · 11 August 2025 · secondary source
- NIST, Center for AI Standards and Innovation (CAISI): CAISI's Assessment of Z.ai's GLM-5.3 Cyber Capabilities · 17 September 2026 · primary source
- NYU Shanghai RITS (Utku Ege Tuluk): GLM-5.3: Z.ai Scales Post-Training, Holds the Weights Back · 17 August 2026 · secondary source
- Mervin Praison (mer.vin): GLM-5.3 Ships Open Weights After Two Weeks as an API-Only Model · 28 August 2026 · secondary source
- 新京报 (Beijing News), 罗亦丹: 智谱GLM-5.3发布,涌现网络安全能力 · 14 August 2026 · secondary source
- Z.ai Security: Z.ai Security Disclosure Ledger (cvd.z.ai) · 3 November 2026 · secondary source
- Z.ai: Z.ai Terms of Use · 14 April 2026 · secondary source
- DoNews (杨亮): 公安部通报35款违法收集个人信息APP 智谱清言、台铃电动在列 · 20 May 2025 · secondary source
- IT之家 (远洋): 官方通报 35 款 App 违法违规收集个人信息:涉多款剪辑 App · 20 May 2025 · secondary source
3 · Transparency
Z.ai discloses weights, training data volume and benchmarks, but its latest models are released without a section on safety and limits.
What is documented
- The technical reports on GLM-4.5 and GLM-5 describe architecture, training, token counts and source categories such as web, code, books and papers3 4
- The weights of the GLM models are published on Hugging Face, as are evaluation tools5 3
- The GLM-4.5 report gives a SafetyBench result and discloses unfairness and bias as the weakest area3
- The FLI describes Z.ai as more transparent than its Chinese peers6
- The terms of use require visible labelling of AI-generated output and prohibit removing AI Seals1
The case against
Gaps on limits and safety
- The GLM-5 report and the GLM-5.3 model card contain no section on safety, limits or usage restrictions4 5
- System prompts and behaviour specification are not disclosed; the FLI grades Information Sharing D6
- The technical implementation of labelling and the registration numbers of the services are not documented1 2
Sources for Transparency (6)
- Z.ai: Z.ai Terms of Use · 14 April 2026 · secondary source
- 国家互联网信息办公室 (CAC): 国家互联网信息办公室关于发布生成式人工智能服务已备案信息的公告 · 2 April 2024 · primary source
- GLM-4.5 Team, Zhipu AI & Tsinghua University (arXiv): GLM-4.5: Agentic, Reasoning, and Coding (ARC) Foundation Models (arXiv:2508.06471v1) · 8 August 2025 · primary source
- GLM-5 Team, Zhipu AI & Tsinghua University (arXiv): GLM-5: from Vibe Coding to Agentic Engineering (arXiv:2602.15763v2) · 24 February 2026 · primary source
- Z.ai (zai-org) auf Hugging Face: Model card zai-org/GLM-5.3 · 27 August 2026 · primary source
- Future of Life Institute: AI Safety Index – Summer 2026 (Digital, 14 Jul) · 14 July 2026 · primary source
4 · Why the harm rule does not apply
The harm rule does not apply: during the period there is no judgment, regulatory decision or fine concerning harm caused by the company's own AI systems. The Entity List is an export-control classification of the company, and the finding on the chat app concerns data collection.
5 · Controversies (3)
- mediumA state testing body found that the 智谱清言 app (version 2.9.6) collected personal data beyond user consent. · 20 May 2025 · abgeschlossen
- mediumThe US Department of Commerce added seven Zhipu entities to the Entity List for supporting China's military modernisation; Zhipu disputes this. · 16 January 2025 · open
- minorOwing to a shortage of computing capacity at the launch of GLM-5, Zhipu restricted access in the GLM Coding Plan, apologised and offered refunds. · 21 February 2026 · unklar
6 · What remains open
- Prospectus of 30 December 2025 with the chapters on risk factors, regulation, litigation and governance
- Annual report 2025 with the corporate governance report
- Frontier safety framework under the Seoul commitment and release criteria for open weights
- System card or safety report on GLM-5.3 and GLM-5.2 including the content of the safety evaluation
- Original notice of the official finding on the 智谱清言 app and the remediation
- Registrations of Zhipu services in the CAC lists of generative AI services
7 · How it was assessed
Cut-off date: 3 November 2026 · 39 searches, 37 fetches.
Primary documents read (10)
- ESG Report 2025 by Knowledge Atlas Technology JSC (S01, two passes)
- Federal Register 90 FR 4617, Entity List (S02, reproduction)
- HKEX announcements dated 2025-12-19 and 2025-12-30 (S28, S29)
- Technical reports GLM-4.5 and GLM-5 (S07, S08)
- Model card GLM-5.3 (S09)
- CAISI assessments GLM-5.2 and GLM-5.3 (S11, S12)
- FLI AI Safety Index Summer 2025 and Summer 2026 (S03, S04)
- Frontier AI Safety Commitments (S05)
- Z.ai Terms of Use and Security Disclosure Ledger (S21, S20)
- CAC announcement 2024-04-02 (S30)
Research of 3 November 2026 under Globalance AI Footprint rulebook version 0