Broad AI governance and testing, open questions on military and government clients
Overall rating “solid”. No deduction for controversies. The superscript numbers link directly to the source.
1 · Responsibility
Oversight and review programmes are well documented, but the Unit 8200 case shows gaps in how relationships with security-agency clients are governed.
What is documented
- A board committee oversees AI governance and regulatory risks as well as the Responsible AI programmes, and the full board oversees AI strategy risk1 2
- Operationally there are a Responsible AI Council, the Office of Responsible AI and a Chief Responsible AI Officer3 4
- The Responsible AI Standard revised in 2026 contains its own chapter on procured third-party AI4
- In the reporting period there were around 2,500 pre-release reviews of generative AI and more than 450 sensitive-use reviews4
- Microsoft 365 Copilot is certified to ISO/IEC 4200110 4
The case against
Military and government clients
- According to reports, Unit 8200 used Azure for intercepted phone calls of Palestinians; Microsoft's review confirmed parts of this, and remediation is only partly documented6 7 8
- Leaked documents show ICE tripled its data storage on Azure; Microsoft denies mass surveillance9
- A complaint by the ICCL to the Irish Data Protection Commission is pending12
Gaps in accountability
- The frequency and content of AI reporting to the board committee are not disclosed
- There are no publicly documented remedy cases from external grievance channels for affected people
Sources for Responsibility (12)
- Microsoft Corporation / SEC: Form DEF 14A – Proxy Statement 2025 · 5 December 2025 · primary source
- Microsoft Corporation / SEC: Form DEF 14A – Proxy Statement 2026 · 22 October 2026 · primary source
- Microsoft: 2025 Responsible AI Transparency Report · 20 June 2025 · primary source
- Microsoft: Responsible AI Transparency Report 2026 · 1 September 2026 · primary source
- Microsoft: Frontier Governance Framework (February 2026 version) · 3 November 2026 · primary source
- Capacity Media: Microsoft cuts cloud and AI services to Israeli military unit · 26 September 2025 · secondary source
- DatacenterDynamics: Microsoft wraps up investigation into Israeli military's use of its cloud platform · 4 June 2026 · secondary source
- Ynetnews: Microsoft changes its human rights policy following review of IDF's use of its cloud technology · 7 June 2026 · secondary source
- TechRepublic: Microsoft Denies 'Mass Surveillance' as ICE Triples Down on Azure Usage · 19 February 2026 · secondary source
- Microsoft Tech Community: How Microsoft 365 Copilot builds customer trust with responsible AI · 1 March 2025 · primary source
- GeekWire: Human rights proposals win more than 25% of votes at Microsoft shareholder meeting · 5 December 2025 · secondary source
- Law Society Gazette (Ireland): ICCL files Gaza complaint against Microsoft · 5 December 2025 · secondary source
2 · Safety
Red-teaming, external partners and a frontier framework are documented, set against an official finding on security culture and recurring Copilot flaws.
What is documented
- The AI Red Team carried out more than 100 operations in 2025; an external red-team alliance includes 18 universities, alongside cooperation with CAISI and UK AISI2
- The Frontier Governance Framework sets thresholds, provides for tiered protection of model weights and governs reporting of serious incidents to authorities and customers3
- The Secure Future Initiative reports measurable progress, for example 99.6% phishing-resistant multi-factor authentication6
- Incidents are governed by 13 playbooks; more than 600 AI safety cases were escalated in 20252 1
- The Digital Crimes Unit sued Storm-2139, a network that misused stolen Azure OpenAI access for celebrity deepfakes14 1
The case against
Security flaws and security culture
- In 2024 the Cyber Safety Review Board found that the Storm-0558 intrusion was avoidable and that the corporate culture neglected security7 8
- EchoLeak allowed data leakage from Microsoft 365 Copilot without any user action; the flaw has been fixed and no exploitation is known9
- Copilot Chat summarised emails marked confidential despite protection policies; according to a report, CoSnitch was fixed only more than eight months after it was reported10 11
Harmful outputs and government environments
- According to 404 Media, sexualised deepfakes of Taylor Swift were created with Microsoft Designer; Microsoft then tightened its safeguards15
- One of Microsoft's own engineers reported harmful image outputs from Copilot Designer to the FTC and the board4
- According to ProPublica, China-based engineers maintained US Defense Department cloud systems via “digital escorts”; Microsoft ended the practice13
Sources for Safety (15)
- Microsoft: 2025 Responsible AI Transparency Report · 20 June 2025 · primary source
- Microsoft: Responsible AI Transparency Report 2026 · 1 September 2026 · primary source
- Microsoft: Frontier Governance Framework (February 2026 version) · 3 November 2026 · primary source
- FOX 13 Seattle (Associated Press): Microsoft engineer sounds alarm on AI image-generator to US officials and company's board · 6 March 2024 · secondary source
- Insurance Journal: Estate of Suzanne Adams sues OpenAI, Sam Altman and Microsoft · 12 December 2025 · secondary source
- Microsoft Security Blog: Securing our future: November 2025 progress report on Microsoft's Secure Future Initiative · 10 November 2025 · primary source
- U.S. Department of Homeland Security: Cyber Safety Review Board Releases Report on Microsoft Online Exchange Incident from Summer 2023 · 2 April 2024 · primary source
- CISA: Cyber Safety Review Board Releases Report on Microsoft Online Exchange Incident from Summer 2023 · 3 November 2026 · primary source
- The Hacker News: Zero-Click AI Vulnerability Exposes Microsoft 365 Copilot Data Without User Interaction · 12 June 2025 · primary source
- The Register: Microsoft Copilot data loss prevention bug (Issue No. 1226324) · 18 February 2026 · secondary source
- OECD.AI – AI Incidents Monitor: Microsoft Copilot Vulnerability Exposed Sensitive Data via AI Prompt Injection · 18 August 2026 · secondary source
- Help Net Security: Microsoft delays Windows Recall rollout, more security testing needed · 14 June 2024 · secondary source
- ProPublica: Microsoft's use of China-based engineers and «digital escorts» for Defense Department cloud · 18 July 2025 · primary source
- The Register: Microsoft names and shames 4 alleged Azure OpenAI abusers (Storm-2139) · 28 February 2025 · secondary source
- 404 Media: Microsoft Closes Loophole That Created AI Porn of Taylor Swift · 29 January 2024 · primary source
3 · Transparency
Microsoft discloses intended uses, limitations and model information in detail, but did not take part in the independent 2025 transparency index.
What is documented
- The Transparency Note for Azure OpenAI describes intended uses, excluded uses such as high-risk decisions, bias, hallucinations and limits of agentic features5
- Since 2026 there have been a model card and a training data summary for MAI-Image-1 under the EU Code of Practice, as well as 20 application and platform cards2
- Under California law AB 2013, summaries are available for MAI-Image-1 and Phi-4-Reasoning2
- Since 2023 Bing Image Creator, Copilot, Designer and Paint have attached C2PA provenance data to content, combined with watermarks1 2
The case against
No independent comparison
- Microsoft declined to take part in the 2025 Foundation Model Transparency Index7
- The Transparency Note for Microsoft 365 Copilot and the model cards of the MAI models were not reviewed in the original
Proceedings on disclosure and training data
Sources for Transparency (9)
- Microsoft: 2025 Responsible AI Transparency Report · 20 June 2025 · primary source
- Microsoft: Responsible AI Transparency Report 2026 · 1 September 2026 · primary source
- iTnews: ACCC accuses Microsoft of misleading 2.7 million Australians over M365 fees · 27 October 2025 · secondary source
- iTnews: Microsoft in damage control over Copilot bundling bungle · 6 November 2025 · secondary source
- Microsoft Learn: Transparency Note for Azure OpenAI · 21 September 2026 · primary source
- Microsoft Tech Community: How Microsoft 365 Copilot builds customer trust with responsible AI · 1 March 2025 · primary source
- arXiv (Stanford CRFM et al.): The 2025 Foundation Model Transparency Index (arXiv:2512.10169v1) · 3 November 2026 · primary source
- Stanford CRFM: The Foundation Model Transparency Index (December 2025) · 3 November 2026 · primary source
- Axios: NYT case against OpenAI and Microsoft can advance · 1 April 2025 · secondary source
4 · Why the harm rule does not apply
The harm rule does not apply: the official findings within the time window do not concern harm caused by AI, and there is no pattern of serious, well-documented AI incidents. The ACCC lawsuit and the ICCL complaint have not yet been decided.
5 · Controversies (18)
- mediumLeaked documents show tripled Azure use by ICE; Microsoft denies mass surveillance. · 19 February 2026 · unklar
- mediumThe estate of Suzanne Adams is suing OpenAI and Microsoft after a homicide-suicide following ChatGPT conversations. · 11 December 2025 · open
- mediumAccording to the ACCC lawsuit, Microsoft misled around 2.7 million customers in Australia about the subscription option without Copilot. · 27 October 2025 · open
- mediumAccording to reports, Unit 8200 of the Israeli military used Azure for intercepted phone calls; Microsoft's review confirmed parts. · 25 September 2025 · unklar
- mediumAccording to ProPublica, China-based engineers maintained US Defense Department cloud systems via “digital escorts”. · 18 July 2025 · unklar
- mediumLawsuit by the New York Times against OpenAI and Microsoft for copyright infringement through training and outputs. · 1 April 2025 · open
Further controversies (12)
- mediumThe Cyber Safety Review Board found that the Storm-0558 intrusion was avoidable and that the culture neglected security. · 2 April 2024 · abgeschlossen
- mediumA Microsoft engineer told the FTC and the board that Copilot Designer generated violent and sexualised images. · 6 March 2024 · unklar
- mediumAccording to 404 Media, sexualised deepfakes of Taylor Swift were created with Microsoft Designer and spread virally. · 25 January 2024 · unklar
- minorCoSnitch: a prompt-injection flaw in Copilot allowed data leakage; according to a report, the patch came more than eight months after it was reported. · 18 August 2026 · unklar
- minorSecurities class action by a pension fund: Microsoft allegedly overstated the success and use of Copilot. · 12 June 2026 · open
- minorCopilot Chat summarised emails marked confidential despite protection policies; the bug has been fixed. · 21 January 2026 · unklar
- minorShareholder proposals on human rights, data centres and AI data use were rejected. · 5 December 2025 · abgewiesen
- minorThe Austrian data protection authority found a breach of the right of access in Microsoft 365 Education. · 13 October 2025 · abgeschlossen
- minorA report by the UN Special Rapporteur lists Microsoft among companies profiting from the occupation. · 3 July 2025 · unklar
- minorAuthors' lawsuit against Microsoft over the alleged use of pirated books for model training. · 25 June 2025 · unklar
- minorEchoLeak: a zero-click flaw in Microsoft 365 Copilot enabled data leakage; no exploitation is known. · 11 June 2025 · unklar
- minorSecurity researchers showed protection flaws in Windows Recall; Microsoft delayed the launch and made improvements. · 7 June 2024 · unklar
6 · What remains open
- Independent finding on the AI dimension of use by Unit 8200 and ICE
- The external review by Covington & Burling of the Israeli Ministry of Defence's use of Azure, in the original
- Court records of the lawsuits by the New York Times, the estate of Suzanne Adams and the ACCC; so far only secondary sources
- Independent audit of the security and Responsible AI metrics
- Link between AI or security targets and executive pay
- Outcome of the shareholder vote on 8 December 2026 on the human rights proposals
7 · How it was assessed
Cut-off date: 3 November 2026 · 0 searches, 39 fetches.
Primary documents read (11)
- Form 10-K FY2025 (S09, two retrievals)
- Form 10-K FY2026 (S10, first 100,000 characters)
- Proxy Statement DEF 14A 2025 (S07, first 100,000 characters)
- Proxy Statement DEF 14A 2026 (S08, first 100,000 characters)
- Responsible AI Transparency Report 2025 (S03, first 100,000 characters)
- Responsible AI Transparency Report 2026 (S04, first 100,000 characters)
- Frontier Governance Framework February 2026 (S06)
- Transparency Note for Azure OpenAI (S35)
- Secure Future Initiative Progress Report November 2025 (S18)
- DHS/CISA notice on the CSRB report (S37, S36)
- Foundation Model Transparency Index 2025 (S23, S24)
Research of 3 November 2026 under Globalance AI Footprint rulebook version 0